Computational Biology

  • 90 511 10
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up

Computational Biology

Chapman & Hall/CRC Mathematical and Series A STATISTICAL MECHANICS PERSPECTIVE CHAPMAN & HALL/CRC Mathematical and

1,401 234 30MB

Pages 276 Page size 432 x 648 pts Year 2008

Report DMCA / Copyright


Recommend Papers

File loading please wait...
Citation preview

Chapman & Hall/CRC Mathematical and Computational Biology Series



CHAPMAN & HALL/CRC Mathematical and Computational Biology Series Aims and scope: This series aims to capture new developments and summarize what is known over the whole spectrum of mathematical and computational biology and medicine. It seeks to encourage the integration of mathematical, statistical and computational methods into biology by publishing a broad range of textbooks, reference works and handbooks. The titles included in the series are meant to appeal to students, researchers and professionals in the mathematical, statistical and computational sciences, fundamental biology and bioengineering, as well as interdisciplinary researchers involved in the field. The inclusion of concrete examples and applications, and programming techniques and examples, is highly encouraged.

Series Editors Alison M. Etheridge Department of Statistics University of Oxford Louis J. Gross Department of Ecology and Evolutionary Biology University of Tennessee Suzanne Lenhart Department of Mathematics University of Tennessee Philip K. Maini Mathematical Institute University of Oxford Shoba Ranganathan Research Institute of Biotechnology Macquarie University Hershel M. Safer Weizmann Institute of Science Bioinformatics & Bio Computing Eberhard O. Voit The Wallace H. Couter Department of Biomedical Engineering Georgia Tech and Emory University

Proposals for the series should be submitted to one of the series editors above or directly to: CRC Press, Taylor & Francis Group 24-25 Blades Court Deodar Road London SW15 2NU UK

Published Titles Cancer Modelling and Simulation Luigi Preziosi Computational Biology: A Statistical Mechanics Perspective Ralf Blossey Computational Neuroscience: A Comprehensive Approach Jianfeng Feng Data Analysis Tools for DNA Microarrays Sorin Draghici Differential Equations and Mathematical Biology D.S. Jones and B.D. Sleeman Exactly Solvable Models of Biological Invasion Sergei V. Petrovskii and Lian-Bai Li An Introduction to Systems Biology: Design Principles of Biological Circuits Uri Alon Knowledge Discovery in Proteomics Igor Jurisica and Dennis Wigle Modeling and Simulation of Capsules and Biological Cells C. Pozrikidis Normal Mode Analysis: Theory and Applications to Biological and Chemical Systems Qiang Cui and Ivet Bahar Stochastic Modelling for Systems Biology Darren J. Wilkinson The Ten Most Wanted Solutions in Protein Bioinformatics Anna Tramontano

Chapman & Hall/CRC Mathematical and Computational Biology Series




Boca Raton London New York

Chapman & Hall/CRC is an imprint of the Taylor & Francis Group, an informa business

C5564_Discl Page 1 Thursday, April 20, 2006 9:59 AM

Published in 2006 by Chapman & Hall/CRC Taylor & Francis Group 6000 Broken Sound Parkway NW, Suite 300 Boca Raton, FL 33487-2742 © 2006 by Taylor & Francis Group, LLC Chapman & Hall/CRC is an imprint of Taylor & Francis Group No claim to original U.S. Government works Printed in the United States of America on acid-free paper 10 9 8 7 6 5 4 3 2 1 International Standard Book Number-10: 1-58488-556-4 (Hardcover) International Standard Book Number-13: 978-0978-58488-556-6 (Hardcover) This book contains information obtained from authentic and highly regarded sources. Reprinted material is quoted with permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts have been made to publish reliable data and information, but the author and the publisher cannot assume responsibility for the validity of all materials or for the consequences of their use. No part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, please access ( or contact the Copyright Clearance Center, Inc. (CCC) 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe.

Visit the Taylor & Francis Web site at Taylor & Francis Group is the Academic Division of Informa plc.

and the CRC Press Web site at

To SyMaNi and Mum

Preface This is not a biophysics book. The readers of this book will indeed find a number of topics in it which commonly classify as biophysics; an example is the discussion of the electrostatic properties of biomolecules. But the book’s ambition is different from being yet another introduction into biophysics. I therefore like to explain my motivation for the selection of topics I made. The title of the book establishes a link between computational biology on the one side, and statistical mechanics on the other. Computational biology is the name of a new discipline. It is probably fair to say that it is, in fact, an emerging one, since a new name is not sufficient to establish a new discipline. Quantitative methods have been applied to biological problems for many years; this has led to a vast number of different subdisciplines of established fields of science: there is a mathematical biology, a biomathematics, a biostatistics, a bioinformatics, a biophysics, a theoretical biology, a quantitative biology... this list is certainly not exhaustive. All these subdisciplines emerged out of the existing fields and developed by an act of transfer: the use of a method or a mathematical approach within a new context, its application to new problems. One may expect that at some point in the future all these different subdisciplines may merge into a common discipline. For the time being and for the lack of a definite name, let us call this discipline computational biology. This book wants to contribute a particular element to this field; the use of statistical mechanics methods for the modelling of the properties of biological systems. Statistical physics is the scientific discipline which was developed in order to understand the properties of matter composed of many particles. Traditionally, it has been applied to non-living matter: gases, liquids, and solids. Meanwhile, it is increasingly applied to what is nowadays called ‘soft matter’ which encompasses complex objects like colloids, membranes and biomolecules, hence objects which do not clearly fall into any one of the classic categories. Statistical mechanics methods are therefore indeed essential in their application to biophysical problems, since they are needed to understand the static and dynamic properties of biomolecules, complex molecular machines and even whole cell behaviour. But there is a second aspect for which these methods can prove important, and this relates to the information content of the biological systems. Biology

is built on recognition processes: DNA strands have to recognize each other, proteins have to identify DNA binding sites etc. In bioinformatics, these recognition problems are commonly modelled as pattern recognition problems: this mapping is the basis of the enormous success of the field of modern genomics. Moving beyond genomics, however, to the biology of whole systems, it becomes increasingly clear that an understanding of the physical properties of biological systems becomes more and more important for processes involving biological information content: DNA is not merely a string spelled out in a four-letter alphabet, but it is also an elastic string. There is biological information contained in its structure and its dynamics. Biology employs physical mechanisms to organize its information processes. This becomes particularly evident, as we begin to understand, in the properties of chromatin, the DNA-protein complex making up the chromosomes in the nuclei of eukaryotic cells (a topic which is discussed in the book). Or, in the particle interaction networks upon which the cellular machinery relies. This book is placed at just this interface: between biological recognition on the one hand, and the statistical physics methods that can be employed to understand the underlying mechanisms. The first part of the book gives a concise introduction into the main concepts of statistical mechanics, equilibrium and nonequilibrium. The exposition is introductory in the choice of topics addressed, but still mathematically challenging. Whenever possible, I have to tried to illustrate the methods with simple examples. The second part of the book is devoted to biomolecules, to DNA, RNA, proteins, and chromatin - i.e., the progression of topics follows more or less what is commonly known as the central dogma of molecular biology. In this part, mostly equilibrium statistical mechanics is needed. The concern here is to understand and model the processes of base-pair recognition and (supra-) molecular structure formation. The third part of the book is devoted to biological networks. Here, both equilibrium and nonequilibrium concepts introduced in the first part are used. The presentation covers several of the nonequilibrium statistical physics approaches described in Chapter 2 of Part I, and illustrates them on biologically motivated and relevant model systems. Throughout the book, Exercises and Tasks are scattered, most of them in the first part. They are intended to motivate the readers to participate actively in the topics of the book. The distinction between Exercises and Tasks is the following: Exercises should be done in order to verify that the concept

that was introduced has been understood. Tasks are more ambitious and usually require either a more involved calculation or an additional idea to obtain the answer. A final technical note: the book is complemented by a detailed index. Keywords listed are marked in italics throughout the text. In the course of shaping the idea of this book and writing it, I profited from discussions with many colleagues. I am especially thankful to Arndt Benecke, Dennis Bray, Martin Brinkmann, Luca Cardelli, Enrico Carlon, Avi Halperin, Andreas Hildebrandt, Martin Howard, Oliver Kohlbacher, Hans Meinhardt, Thomas Lengauer, Hans-Peter Lenhof, Annick Lesne, Ralf Metzler, Johan Paulsson, Andrew Phillips, Wilson Poon, Helmut Schiessel, Bernard Vandenbunder, Jean-Marc Victor, Pieter Rein ten Wolde, Edouard Yeramian. Finally, I particularly like to thank Andreas Hildebrandt and Martin Howard for their detailed and helpful remarks on an early version of the manuscript. ´ I gratefully acknowledge the hospitality of the Institut d’Electronique, de Micro´electronique et de Nanotechnologie (IEMN) in Villeneuve d’Ascq and ´ the Institut des Hautes Etudes in Bures-sur-Yvette, where parts of this book were written.



Statistical Mechanics


1 Equilibrium Statistical Mechanics 1.1 Z: The partition function . . . . . . . . . . . . . . . . . . . . 1.2 Relation to thermodynamics . . . . . . . . . . . . . . . . . . 1.3 Computing Z . . . . . . . . . . . . . . . . . . . . . . . . . . .

1 1 5 11

2 Nonequilibrium Statistical Mechanics 2.1 Stochastic processes . . . . . . . . . . . . . . . . . . . 2.2 The master equation . . . . . . . . . . . . . . . . . . 2.3 Fluctuation theorems . . . . . . . . . . . . . . . . . . 2.4 The Fokker-Planck and Langevin equations . . . . . . 2.5 Sequence alignment: a nonequilibrium phase transition

41 42 44 49 56 68


. . . .

. . . . .

. . . . .

. . . . .



1 Molecules, Code and Representation 1.1 DNA and RNA: the building blocks . . . . . . . . . . . . . . 1.2 Representing RNA structure . . . . . . . . . . . . . . . . . . 1.3 Proteins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

83 83 88 92

2 DNA 2.1 Thermal stability of DNA: the melting transition . . . . . . . 2.2 The melting profiles of genomic DNA and cDNA . . . . . . . 2.3 Hybridizing DNA . . . . . . . . . . . . . . . . . . . . . . . .

97 97 108 117

3 RNA 3.1 Computing RNA secondary structure: combinatorics . . . . 3.2 The RNA partition function . . . . . . . . . . . . . . . . . . 3.3 RNA phase behaviour and folding kinetics . . . . . . . . . .

129 129 133 139

4 Proteins 4.1 Proteins: folding . . . . . . . . . . 4.2 Proteins: docking . . . . . . . . . 4.3 Electrostatics . . . . . . . . . . . . 4.4 Chromatin . . . . . . . . . . . . . 4.5 Bacterial chemotaxis: cooperativity

149 149 156 158 167 173

. . . . . . . . . . . . once

. . . . . . . . . . . . . . . . more

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .





1 Network Dynamics I: Deterministic 1.1 Deterministic dynamics: λ-repressor expression . . . . . . . . 1.2 The Turing insight . . . . . . . . . . . . . . . . . . . . . . . . 1.3 The Min system . . . . . . . . . . . . . . . . . . . . . . . . .

185 186 194 197

2 Network Dynamics II: Fluctuations 2.1 Noise in signalling . . . . . . . . . . 2.2 Stochastic cascades . . . . . . . . . 2.3 Stochastic focusing . . . . . . . . . 2.4 Fluctuating gene network dynamics 2.5 Extrinsic vs. intrinsic noise . . . . .

. . . . .

207 207 210 215 217 222

3 Networks: Structure 3.1 Networks as graphs . . . . . . . . . . . . . . . . . . . . . . . 3.2 Probability generating functions and network characteristics 3.3 Statistical mechanics of networks . . . . . . . . . . . . . . . . 3.4 Network growth . . . . . . . . . . . . . . . . . . . . . . . . .

229 229 232 235 237



. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

Part I

Statistical Mechanics

Chapter 1 Equilibrium Statistical Mechanics

Apr`es tout, vous ˆetes ici, c’est l’essentiel! Nous allons faire du bon travail ensemble [...]! D’ici peu, le monde ´ebloui decouvrira la puissance du grand Z!

Franquin, Z comme Zorglub (1961)


Z: The partition function

This section introduces the basic physical concepts and mathematical quantities needed to describe systems composed of many particles, when only a statistical description remains possible. Already for an apparently trivial system such as a gas of identical atoms, composed of N ∼ 1023 particles/mole (Avogadro’s number), the description of each particle’s trajectory is illusory and not even desirable, and hence only a statistical approach feasible. In biology, as we will see, things are more complicated: not only is the number of relevant interacting components large, but in addition the components are very often ‘individuals’. Suppose we can characterize our system of interest by a given number of ‘state variables’, whose number can be finite or infinite. An example of such a state variable can be the spatial position of a particle, hence a continuous variable, but in principle it can be any other of its distinguishing characteristics. We call the set of these state variables x, irrespective of their physical nature, and treat x here as a set of discrete variables;1 we call such a collection of variables characterizing the state of a system a microstate. The possible microstates a system can assume, typically under certain constraints, can be subsumed under the name of a statistical ensemble. One has to distinguish the 1 In

the case of continuous variables, sums have to replaced by integrals in an obvious way; we will encounter this later and will pass liberally from one notation to the other.



Computational Biology: A Statistical Mechanics Perspective

microstate the system can assume in each of its realisations from the ultimate macrostate that is to be characterized by a macroscopic physical observable after a suitable averaging procedure over the microstates.2 Example. We give a simple illustration of the concept by considering the conformations of a linear molecule such as DNA. The macrostate is the chain conformation we will most probably observe when we image the molecule under certain experimental conditions, while a microstate is any mechanically possible conformation of the molecule. The collection of all these possible states is what we call the statistical ensemble. Coming back to the conceptual question: what quantity governs the probability P (x) of each state to occur? Probability and information. We call P (x) ≥ 0 a probability; conseP quently, {x} P (x) = 1, where the summation is over all possible states. To each probability we can associate an information measure3 s(x) = − ln P .


For P = 1, s = 0, and for P = 0, s = ∞. Therefore s(x) is a measure for the lack of information. It can be used to define entropy as the average of s(x) in distribution, X X S(P ) = s(x)P (x) = − P (x) ln P (x) . (1.2) {x}


In our terminology entropy can also be called an average ignorance. The information-theoretic unit of entropy is the bit. Example. Again we take as a simple example DNA, which consists of the four bases A, C, G, T (the details of the build-up of DNA are explained in Part II of the book). If the probability of all bases is identical, P (base) = 1/4, and we have S(P ) = 4 × (1/4) × ln2 (4) = 2. The average ignorance is thus 2 bits per nucleotide. In order to interpret this result consider the following: before a base is read by some device (a polymerase transcribing DNA into RNA, see Part III), the information can be represented in binary code as two bits: 11, 10, 01, 00. After reading, the uncertainty has become 0.

2 The

averaging procedure means that we can determine the macroscopic state just by an average over the ensemble of microstates; the resulting ensemble average gives a correct description of the macrostate of the system if the system had sufficient time in its dynamic evolution to sample all its microstates; this condition is called ergodicity. 3 The choice for the logarithm will become clear in the section on thermodynamics. Note that we do not distinguish in our notation between log and ln, the meaning should be clear from the context.

Equilibrium Statistical Mechanics


Maximum entropy. What determines the form of P ? In order to deduce it, we employ the prescription proposed by E. T. Jaynes, 1957. It states that the prior probability distribution maximizes entropy while respecting macroscopic constraints; it should thus yield the maximum average ignorance under those constraints. Consider first the case of equiprobabilities, as we have assumed for our simple DNA example: there is no other constraint. In order to apply Jaynes’ concept, we have P to maximize S(P ) under the only natural constraint of normalization, {x} P (x) = 1. This leads to a variational problem4 δ[S + λ[


P (x) − 1]] = 0



in which δ denotes the P variation of the bracketed term with respect to P ; e.g., one has δS(P ) = − {x} [δP · ln P + P · δ(ln P )], and δ ln P = [d(ln P )/dP ]δP . In eq.(1.3), λ is a Lagrange multiplier, a constant to be determined in the calculation. This variational problem eq.(1.3) is solved by (Exercise!) P (x) = eλ−1 = const. ≡ Ω −1


where Ω is the number of realizations of x. All states are indeed found to be equally probable, with a probability inverse to their number. From the definition of entropy, eq.(1.2), we have the relation S = ln Ω .


The canonical ensemble. We now repeat the above argument and calculation for the so-called canonical ensemble. Here, one additional constraint appears, since we want to characterize the states of the system now by an energy E(x), which we allow to fluctuate with average hEi = E0 . In this case we must maximize S(P ) under two constraints, the normalization condition, and in addition the condition on the average energy X hEi = E(x)P (x) . (1.6) {x}

4 For

those confused by the δ-notation common to variational calculations, the same result is obtained by replacing P → P0 +  × P with a small parameter , and requiring the terms in linear order in  to vanish.


Computational Biology: A Statistical Mechanics Perspective

The resulting variational problem involves two Lagrange parameters, λ and β: X X δ[S + λ[1 − P (x)] + β[hEi − E(x)P (x)]] = {x}




δP (x) · (− ln P − 1 − λ − βE(x)) = 0 ,



with the canonical or Gibbs distribution as a result, Pβ (x) = Zβ−1 e−βE(x) ,


where Zβ is the partition function Zβ =


e−βE(x) .



For dimensional reasons, the Lagrange parameter β must be an inverse energy, which we take as the thermal energy β −1 = kB T


with Boltzmann’s constant kB . We define the free energy as F ≡ −kB T ln Z .


The meaning of this definition will become clear in the following section. Eq.(1.9) is the key quantity in all of the first chapter of this part of the book; we will be mostly concerned with methods how to compute it. Task. Determine Pβ,µ if not only the average energy hEi is taken as a constraint condition, but also the number of particles N is allowed to fluctuate with average hN i. The Lagrange parameter µ is called chemical potential.5 The ensemble governed by the resulting distribution is called the grand canonical ensemble. Equivalence of ensembles. The different ensembles (canonical, grand canonical etc.) are macroscopically equivalent for equilibrium states in the thermodynamic limit of N , V → ∞ with √ n = N/V = const. Deviations from the thermodynamic limit scale as ∼ 1/ N , and are hence negligible since the value of N we usually talk about in statistical mechanics is on the order of Avogadro’s constant.6 This equivalence allows to choose the ensemble based 5 The

chemical potential often confuses: it is a measure for the availability of particles, and plays thus an analogous role as temperature does in providing an energy source. 6 There are exceptions for the equivalence of ensembles, a point we do not pursue here. And, obviously, for systems in which N is very much smaller than Avogadro’s constant, the deviations from the thermodynamic limit can be important.

Equilibrium Statistical Mechanics


on its convenience in a particular application - we will encounter such cases in the second part of the book.


Relation to thermodynamics

The knowledge of the statistical distributions for the different ensembles is sufficient to characterize the macroscopic properties of a physical system in thermal equilibrium. We now make the link explicit between the statistical description and the expressions of macroscopic thermodynamics, for which a number of basic principles can be formulated. The Laws of Thermodynamics. The physical insight of macroscopic thermodynamics is usually summarized in the laws of thermodynamics. They are • Law 0: If two systems A and B are in thermodynamic equilibrium, and B is in equilibrium with C, then A and C are also in equilibrium with each other. The equilibria can be characterized by their mechanical, thermal or chemical properties. • Law 1: Energy E is conserved: dE = dQ + dW


where Q is the heat flowing into the system, and W the work done by the system. If a process is adiabatic, dQ is zero. • Law 2: In a thermally isolated macroscopic system, entropy never decreases. • Law 3: As the temperature of a system tends to absolute zero, entropy reaches a constant. We see that energy and entropy are the key notions of equilibrium thermodynamics. Let us understand them more deeply. To achieve this we first generalize the descriptors we have introduced before and introduce the important notion of an extensive variable. An extensive variable is a variable which scales linearly with system size. Volume is a (trivial) example for such a system property: if one starts with two systems of volume V1 and V2 , the volume of the systems is additive: V = V1 + V2 , and hence scales linearly. The particle number N is another example. Temperature T , however, is not extensive: putting two systems of


Computational Biology: A Statistical Mechanics Perspective

equal temperature T together does not result in a system with temperature 2T . We call the state of a system an equilibrium state if it can be characterized by its (internal) energy E, and a set of extensive parameters X1 , ...Xm . For such a state there exists a function of the extensive parameters, the entropy S (which we obtained before from the Jaynes principle). We call the relation S = S(E, X0 , ..., Xm )


a fundamental equation for the system. S itself is also extensive; furthermore, it is an increasing function7 of E. Consider now the first statement, the notion of the extensivity of S. We can express it mathematically by writing S as a first-order homogeneous function of its extensive parameters S(λE, λX0 , ...λXm ) = λS(E, X0 , ..., Xm ) ,


where λ is a scalar. The second statement about S, the monotonicity property, corresponds to the condition8   ∂S ≥ 0. (1.15) ∂E Xi The monotonicity property of S with respect to E allows its inversion, leading to E = E(S, X0 , ...., Xm )


and E, like S, is a first-order homogeneous function E(λS, λX0 , ...λXm ) = λE(S, X0 , ..., Xm ) .


Some properties of scalar fields. Before introducing the thermodynamic potentials, we need some mathematical concepts for scalar fields. They are stated here just as facts.


is also extensive, but this is actually a subtle point if systems become strongly interacting. We are not concerned with sophisticated problems of this sort here. 8 The equality is restricted to a set of measure zero.

Equilibrium Statistical Mechanics


• If φ = φ(x0 , ..., xm ) is a scalar field of m+1 variables, its total differential is given by m X ∂φ dφ = dxi . (1.18) ∂x i i=0 • If the xi = xi (u, v) with u, v scalar, we can rewrite this expression as dφ =

m m X X ∂φ ∂xi ∂φ ∂xi du + dv . ∂xi ∂u ∂xi ∂v i=0 i=0


• Contour surfaces, for φ = constant, define an implicit functional relationship between the xi since on a contour surface m X ∂φ dxi = 0 . ∂x i i=0


• If all xi except x0 and x1 (e.g.) are held fixed, then 

∂x1 ∂x0

 =− φ,...

∂φ ∂x0

 · x1 ,...


∂x0 ∂x1

 = φ,...

∂x1 ∂x0

∂φ ∂x1

 −1 ,


x0 ,...

 −1 (1.22) φ,...

for φ = const. • For three variables, one has the cyclic rule       ∂x1 ∂x2 ∂x0 · · = −1 , ∂x1 x2 ∂x2 x0 ∂x0 x1


which generalizes to more variables in an obvious way (Exercise). We can now proceed to look at the total differentials of E and S. We assume that E is a function of entropy S, volume V , and particle number N ; the extension to further extensive variables is straightforward. We have       ∂E ∂E ∂E dE = dS + dV + dN (1.24) ∂S V,N ∂V S,N ∂N S,V where the subscripts indicate the variables to be held fixed. We define the following intensive parameters:   ∂E T = (1.25) ∂S V,N


Computational Biology: A Statistical Mechanics Perspective

is temperature;  P =−

∂E ∂V

 (1.26) S,N

is pressure;  µ=

∂E ∂N

 (1.27) S,V

is chemical potential; hence dE = T dS − P dV + µdN .


The intensive parameters are all functions of S, V, N , and the functional relationships in eqs.(1.25) - (1.27) are called equations of state. Exercise. Verify explicitly that temperature T , pressure P and chemical potential are intensive variables, i.e., that they are homogeneous functions of zeroth order. Exercise. Deduce from the fact that E is a first-order homogeneous function of its variables one can obtain the relation (the Euler equation9 ) E = T S − P V + µN .


Hint: differentiate eq.(1.17) with respect to λ and put λ = 1. Given that we have introduced E as a function of entropy S, volume V and particle number N , one may wonder whether it is not possible - and even desirable - to define thermodynamic functions of other variables than the ones chosen. In particular, one may also want to study dependences on intensive rather than extensive variables. This is indeed possible, and the resulting functions are not independent from each other. In fact, one can pass from one to the other via a mathematical transformation we introduce first in a formal manner. Legendre transform. Let Y (X0 , ..., Xm ) be a scalar field of the extensive variables Xj , and the Pj = (∂Y /∂Xj )Xi6=j are the corresponding intensive variables. Then Λ = Y [P0 , ..., Pi ]i≤m ≡ Y −

i X

Xj Pj



9 The

expressions for E show that T S has the dimension of an energy. Hence the information entropy we determined in the first section has to be multiplied by Boltzmann’s constant, kB .

Equilibrium Statistical Mechanics


is the Legendre transform of Y with respect to Xj≤i . The total differential of Λ reads i m X X dΛ = − Xj dPj + Pj dXj (1.31) j=0


so that Λ = Λ(P0 , ..., Pi , Xi+1 , ..., Xm ) is a function of the i + 1 intensive, and m − i extensive variables. After this formal definition we want to apply this concept. This is best done by simple cases, ignoring the physical context for the moment. Legendre transform in one dimension. We first discuss the Legendre transform in one dimension for a function f (x) and its derivative y = f 0 (x) =

df ≡ g(x) . dx


We can understand y = g(x) as a variable transformation from x to y. How can we express f in terms of y? For the function f (x) we can write the equation within each point x along the curve f (x) = xf 0 (x) + b(x) . (1.33) Since x = g −1 (y) where g −1 is the inverse of g, we have b(g −1 (y)) = f (g −1 (y)) − yg −1 (y) ≡ Λ(y) .


This is often written in short form as Λ(y) = f (x) − xy ,


where x is a function of y. We move on to some exercises. Exercise. Compute the Legendre transform of f (x) = ex . Exercise. Give the Legendre transform of the harmonic form with the vector x, u(x) =

1 T (x · A · x) , 2

where A is an invertible (n × n)-matrix.



Computational Biology: A Statistical Mechanics Perspective

We now return to the physical context. For the energy E, the four most common thermodynamic potentials that result from the application of the Legendre transform are the • Helmholtz free energy (T given, canonical ensemble) F (T, V, N ) = E[T ] = E − T S = −P V + µN


dF = −SdT − P dV + µdN


H(S, P, N ) = E[P ] = E + P V = T S + µN


dH = T dS + V dP + µdN


• Enthalpy (P given)

• Gibbs free energy (T, P given) G(T, P, N ) = E[T, P ] = E + P V − T S = µN


dG = −SdT + V dP + µdN


• Grand canonical potential (T, µ given) Φ(T, V, µ) = E[T, µ] = E − T S − µN = −P V


dΦ = −SdT − P dV − N dµ


The thermodynamic potentials all have to be minimized at equilibrium for fixed values of their variables. This ends our brief look into thermodynamics. The message that we want to retain is that • within equilibrium physics there is a well-established body of thermodynamic functions with which the macroscopic properties of a system can be described; • these functions have a rigorous link with each other via the Legendre transform, and with statistical mechanics, since we know how to compute the thermodynamic potentials within this theory. We now return to statistical mechanics, and begin to discuss methods which allow to compute the partition function and the quantities derivable from it.

Equilibrium Statistical Mechanics



Computing Z

This section introduces methods to compute Z and the thermodynamic quantities that can be derived from it. We begin with some technicalities that will be useful later. An obvious first step in the computation of Z is to know how to compute integrals involving the Gaussian distribution. The Gaussian distribution is a generic characteristic of equilibrium states: in thermal equilibrium, a system will be in a state minimizing the corresponding thermodynamic potential, and the fluctuations around this stable state will be Gaussian. Gaussian distribution. The Gaussian probability distribution in the case of one variable −∞ < x < ∞ is given by 1


P (x) = Ce − 2 Ax


where the normalization constant is   1/2 B2 A C= e− 2A . 2π



The parameter A > 0 controls the width of P (x) and, together with B, the peak position. Introducing the average µ and variance σ 2 we find µ1 = −

B , A

σ2 =

1 A


and can thus write the normalized form of the Gaussian or, standard normal distribution (x−µ1 )2 1 e − 2σ2 . P (x) = √ (1.48) 2πσ 2

The multivariate version of the distribution, for i = 1, ..., m random variables x = {xi } is   m m X X 1 Aij xi xj − Bi xi  (1.49) P (x) = C exp − 2 i,j=1 i=1 where A = Aij is a positive definite symmetric matrix. The normalization constant reads in this case as   1 C = (2π) −m/2 (Det A) −1/2 exp − BT · A −1 · B (1.50) 2


Computational Biology: A Statistical Mechanics Perspective

with the matrix inverse A −1 . Based on these results we can now easily write down the mean of each of the xi X hxi i = − (A −1 )ij Bj (1.51) j

and the covariance h(xi − hxi i)(xj − hxj ii = hxi xj i − hxi ihxj i = (A −1 )ij


The inverse of A is the covariance matrix. Its diagonal elements are the variances, the off-diagonals are the co-variances. Characteristic function, moment generating function and cumulants. The characteristic function of a stochastic variable X is defined by Z G(k) ≡ heik i = dxP (x)eikx , (1.53) where the symbol h...i was used to abbreviate the average in distribution. Obviously, G(k) is a Fourier transform of P . It exists for real k and obeys G(0) = 1 , |G(k)| ≤ 1 .


The coefficients µm of its Taylor expansion are the moments ∞ X (ik)m G(k) = µm m! m=0

where µn = hxn i =


dxP (x)xn .



The coefficients κm of the series of its logarithm, ln G(k) =

∞ X (ik)m k=1




are the cumulants. They are combinations of the moments, the lowest of which are κ1 = µ1 , κ2 = µ2 − µ21 = σ 2 . (1.58) For the Gaussian distribution, all higher cumulants vanish. Higher moments can thus serve to characterize more complex distributions - this is useful since

Equilibrium Statistical Mechanics


not in all cases full distributions are obtainable. Exercise. Compute the cumulants for the Poisson distribution pn =

an −n e n!


over the integers and zero, n = 0, 1, 2, .... We have reached a point where we want to see the machinery of statistical mechanics in action. We will now apply it to two standard model systems of statistical mechanics. The first is the Ising model, probably the most famous model ever formulated in statistical mechanics. Originally invented to describe an uniaxial ferromagnet and its phase transition from the paramagnetic (non-magnetic) to the ferromagnetic phase, it has found innumerable applications all over the fields of physics, chemistry and biology. The second model we will introduce is that of the mechanics of a polymer. Again think of DNA, but now on a different length scale, beyond that of the base pairs. The molecule is considered as an elastic string which resists bending in space, but which can also be pulled along its axis. We begin by a discussion of the one-dimensional Ising model and its solution. The 1d-Ising model. The Ising model on a one-dimensional lattice is defined by the energy10 H=−

X KX sn sn+1 − h sn 2 n n


where the sn = ±1 are called the ‘spins’ - local magnetic moments - of a ferromagnet.11 Within the model, the parameter K describes a coupling between neighbouring local spins; for K > 0 they will prefer to be in the same state since the energy favors that state. The parameter h in the second term defines an external coupling affecting each of the spins; for the ferromagnetic system for which the Ising model was originally conceived, this contribution represents an applied magnetic field which, depending on its sign, can favour 10 In

statistical mechanics jargon, the Hamiltonian. ferromagnet is usually a metal which, below a certain temperature, has a permanent magnetic moment; the origin of this permanent moment lies in the microscopic magnetic moments produced by the electron shells inside the metal. In the model description we employ, we summarize these atomistic effects by an elementary spin. This is sufficient if we are only interested in the ordering phenomenon between a paramagnetic (non-magnetic) and a ferromagnetic phase. A true microscopic theory of ferromagnetism has to be a quantum theory. 11 A


Computational Biology: A Statistical Mechanics Perspective

either the ‘up’ or the ‘down’ state with all spins following the field direction. Since we will use the Ising model in a biological context, it is clear that the original interpretation of the parameters is largely immaterial for us: for us, sn is just a two-valued variable, and can be re-interpreted at will. From a biological point of view we can, e.g., consider the Ising model in one spatial dimension as a chain of objects which can be in either of two states. A simple biological example to which this model then applies can be a DNA molecule. Within the Ising model we can define microstates that distinguish each other by the different numbers of spins up - interpreted as ‘base pairs bound’ and spins down, interpreted as ‘base pairs unbound’. Hence we can, e.g., address the question of the binding of two DNA strands, and the Ising model becomes a first - but ultimately too crude - model for the physics of a DNA double strand. This re-interpretability of the model is a big advantage of statistical mechanics: indeed, very often physically quite different situations turn out to fall into the class of just one type of model. This explains the success of the Ising model in so many applications - having a system with just two alternatives is the simplest one can have. Plugging the Hamiltonian of the Ising model into the partition function in the canonical ensemble, we obtain Z=


e −H =

{sn }





sn sn+1 +βh





{sn }

with J ≡ βK/2. In the following we leave out the field h; its inclusion in the calculation is left as a Task. Now we have to compute Z, and we will learn a first method to do it. The transfer-matrix solution of the 1d-Ising model. The partition function can be rewritten as Z=


e Js0 s1 · e Js1 s2 · ... · e JsN −1 s0


{sn }

where a periodic boundary condition is assumed, i.e., the linear chain is joined at its ends. Since each spin has two orientations, there are four configurations for two neighbouring spins, two of which are degenerate in energy. This suggests to introduce a matrix representation with a transfer matrix

Equilibrium Statistical Mechanics

 TJ =

eJ e−J e−J eJ



which is a symmetric matrix, to be diagonalized by an orthogonal matrix O. Since this matrix applies to every neighbouring pair of spins we can express the partition function as (Exercise) Z = T r[TJN ] = T r[(OTJ O−1 )N ]


where the symbol T r stands for ‘Trace’, the sum of the diagonal entries of the matrix. The orthogonal matrix O to diagonalize TJ is given by  O=

0 1 −1 0


and we find for the diagonalized matrix the result




cosh J 0 0 sinh J


so that Z = 2N [coshN J + sinhN J] .


Since in statistical mechanics we are interested in the thermodynamic limit which in this model amounts to consider N → ∞, one finally finds the expression of the partition function Z = 2N coshN J .


Show this as an Exercise. Eq.(1.68) is the first partition function we have calculated explicitly, and we can now calculate physical quantities from it. Let us begin with the internal energy E; it is found to be E = Z −1

X {sn }

He −βH = Z −1

∂Z = −N K tanh(βK/2) . ∂β



Computational Biology: A Statistical Mechanics Perspective

A second, more complex example is the two-point correlation function hsi sj i for |i − j| = r. It is given by X hsi sj i = Z −1 si sj e−βH . (1.70) {sn }

In order to apply the transfer matrix method we first have to find a suitable way to represent it in terms of TJ . One can convince oneself that the expression to write down is σ3 TJr σ3 TJN −r (1.71) where σ3 is the matrix  σ3 =

1 0 0 −1



This expression states that between two well-defined spin states along the chain the transfer matrix has to propagate r times between the spins, and then again N − r times along the other side of the closed chain - remember we maintain the periodic boundary condition. With this we can now write hsi sj i = Z −1 T r[σ3 TJr σ3 TJN −r ] = Z −1 T r[(Oσ3 O−1 ) · (OTJ O−1 )r · (Oσ3 O−1 ) · (OTJ O−1 )N −r ] = 2N Z −1 (coshr J sinhN −r J + sinhr J coshN −r J) = tanh N −r J + tanhr J


where Oσ3 O−1 =

0 1 1 0



With N  r and in the thermodynamic limit we obtain the expression hsi sj i = tanhr J = e −|i−j|/ξ


ξ ≡ | ln(tanh J)| −1


where we defined

as the correlation length. For the 1d-Ising model ξ diverges in the limit T → 0,

Equilibrium Statistical Mechanics


when all spins align either up or down (i.e., the system adopts a homogeneous, oriented state). It will turn out that there is no phase transition in the onedimensional Ising model at a finite temperature (T > 0) between a state of finite magnetization - the oriented state in which all spins point collectively either up or down - and a state without a net magnetization, in which the spins point randomly either up or down with a vanishing average. For this simple example, we are essentially done and the reader can try to extend the model to include the field h, or calculate other thermodynamic quantities to get a feeling for the systems’ properties. But we do not yet want to stop here and take at little deeper look into this system. We have computed thermodynamic quantities in the thermodynamic limit N → ∞. Let us suppose we don’t do that - in fact, in some applications of statistical mechanics in biology, this limit may be difficult to reach, and effects of finite system size may be important. Let us therefore study how the properties of the one-dimensional Ising model are affected upon a change of N . We start from the expression Z[N, J] = T r[TJN ] = T r[(OTJ O−1 )N ]


where OTJ O−1 = eJ · 1 + e−J σ3 ; the N -dependence is now explicit. Suppose we, in a first step, double the lattice size from N to 2N . With (OTJ O−1 )2 = 2(cosh 2J · 1 + σ3 ) = COT J O−1 where

√ C = 2 cosh 2J ,


1 ln(cosh 2J) 2



we arrive at the relation Z[2N, J] = C N Z[N, J] .


The equations (1.79), (1.80) can be read as generators of a flow in the coupling constant J. If we were to iterate the mapping J → J starting from a value J = 1, the iterated map will converge to zero (Exercise). What can we learn from this observation? By these very simple considerations we have in fact found the two fixed points of a renormalization group transformation of the 1d-Ising model - without noticing that such a thing exists. The first fixed point is J = 0, which corresponds to T = ∞. Physically, the system is then in the fully disordered phase, in which the spins point randomly either up or down. The second fixed point is J = ∞ corresponding to T = 0, and this is the fully ordered spin


Computational Biology: A Statistical Mechanics Perspective

state which only exists at zero temperature. Indeed, there is no phase with a finite magnetization M ≡ hsi i 6= 0 for 0 < T ≤ ∞, as we mentioned before. Although we have of course not made a systematic approach to renormalization, there is already something to learn. Suppose we consider the system at a given temperature, and at a given size N . The spins inside the system will have a correlation length ξ over which their orientation (up or down) is correlated. If we now double the system size - what does the correlation do: will it grow stronger or not? It is characteristic of a system right at a phase transition that the correlation length is infinite. The system behaves in a collective (or cooperative) fashion. If the temperature I choose is just that critical temperature, all the steps of increasing system size will not affect the correlation length, and it will stay infinite. For the 1d-Ising model, this is just the case at T = 0. By contrast, if I deviate ever so slightly from that critical temperature, the step of increasing system size will tend to reduce the correlation length, and if this procedure is repeated over and over again12 the system will move away to a system of uncorrelated spins, hence a disordered phase. This whole approach, in its systematic version, allows to compute the properties of all equilibrium phase transitions in a unified way. After this digression on the idea behind the renormalization group, we turn to another approach to compute the partition function of the 1d-Ising model: the computation by recursion. This will turn out to be a very important technique in the second part of the book (Part II: Biomolecules), since recursion methods are at the heart of dynamic programming. Solution of the 1d-Ising model by recursion. The analytic solution of the 1d-Ising model is so easy to obtain since the coupling between two spins is identical. This is not necessarily the case, and we will indeed see in Part II that for the most relevant biological applications this simplifying assumption can not be made. So let’s now make things a little more complicated. For the 1d-Ising model with a neighbour-dependent coupling J(n), we write13 βH = −

N −1 X n=1

J(n)s(n)s(n + 1) −





where we now have also included a neighbour-dependent field H(n) = βh(n). This variant of the Ising-model has been applied to problems in statistical genetics, as suggested by J. Majewski et al., 2001, in a study of epistasis 12 Hence 13 With

the notion of a renormalization group; technically, it is a semi-group. a slight change of notation which should be obvious.

Equilibrium Statistical Mechanics


(i.e., gene-gene interaction). In this interpretation, the spins on the lattice are identified with different genes placed along the DNA molecule rather than the individual base pairs. As before, the total number of states in the model is 2N , hence exponential in N , and the computation to be performed is seemingly exponential in N since we now cannot use the transfer matrix trick anymore due to the neighbour-dependence of couplings. Using a recursion technique, however, we can achieve a computation in linear ‘time’ N . For the construction of the recursion relation we write ZN = ZN + + ZN −


where ZN + refers to the partition function of the chain (which is left open in this case - in contrast to our computation before) in which the last spin at site N points up; the interpretation of ZN − is evident. One finds (Exercise) ZN + = eH(N ) [Z(N −1)+ eJ(N −1) + Z(N −1)− e−J(N −1) ]


and ZN − = e−H(N ) [Z(N −1)+ e−J(N −1) + Z(N −1)− eJ(N −1) ] .


The partition function ZN can now be computed recursively from these expressions. This concludes the discussion of the Ising model in one spatial dimension. Although we saw some calculational approaches at work, we can be a little bit dissatisfied since there is no phase transition between an ordered and a disordered state in this model at a finite temperature. So we are asked to generalize, and one way to do this is to pass on to higher dimensions.14 Since calculations in higher dimensions are more complicated (and in fact, they will become far too complicated for the ambitions of this text) we will try to simplify in another way. This leads us to what is called the mean-field solution of the Ising model. In this approach, spatial dimension will first play no role. Mean-field solution. For the discussion of the mean-field approximation we place the spins on the edges of a hypercube in d space dimensions, and 14 With

this generalization we lose for the moment the biological interpretation of the Ising model for the binding of a DNA double strand. But this doesn’t matter: here we are mainly interested in the problem of the calculation of partition functions; the more specific use of such models within a given biological context follows in the subsequent parts of the book.


Computational Biology: A Statistical Mechanics Perspective

write the energy as H=−

1X J(x, y)s(x)s(y) 2 x,y


where x = a xi ei with the lattice constant a and ei , i = 1, ..., d are the unit vectors on the lattice. J(x, y) is taken as a symmetric matrix. Again we need a technical concept first. Hubbard-Stratonovich transformation. Consider the Gaussian integral relation   r   Z ∞ β 2 β −1 2 β exp Js = dφ exp − J φ + βφs . (1.86) 2 2πJ −∞ 2 It can be understood as a linearization of the argument of the exponential (i.e., the variable s). The price to pay for this operation is the introduction of an auxiliary integration variable. We generalize this step to the weight exp


! βX J(x, y)s(x)s(y) = 2 x,y


# X β X −1 dφ(x) exp − φ(x)s(x) J (x, y)φ(x)φ(y) + β 2 x,y −∞ x

YZ x


where we now have introduced a local order parameter field φ(x) at each lattice site. Note that the expression is meaningful as long as the matrix is semi-definite and positive, which is not the case for the Ising model since J(x, x) = 0; the formula thus has to be considered a formal one.15 The partition sum over the spins s can be performed exactly with the Hubbard-Stratonovich transformation, with the result Z Z=


# X β X −1 Dφ exp − J φ(x)φ(y) + ln(cosh[βφ(x)]) 2 x,y x


R RQ where the integral measure is Dφ ≡ x dφ(x). The ln(cosh(..))-term in the argument of the exponential function arises from the summation over the 15 Further,

we have ignored the integration amplitude.

Equilibrium Statistical Mechanics


spins s(x) which is now easy to do - that was just the idea behind the linearization in the first place. We now have the expressions in place to perfom the mean-field approximation. It corresponds to taking the saddle-point of the integrand, which is determined by the stationary value of the argument of the exponential. This idea is easily explained for a function of one variable; it is also known as the method of steepest descent.16 Suppose the integral we want to compute is Z

I = lim

N →∞

dx e−N f (x)



where we have introduced an explicit parameter N ; frequently such a parameter can be defined, if often only on formal grounds. Suppose the function f has a global minimum at a value x = x0 , well-separated from possibly other minima. Upon Taylor expansion, f fulfills 1 f (x) ≈ f (x0 ) + f 00 (x0 )(x − x0 )2 + O(x3 ) , 2


since we expand around an extremum for which f 0 (x0 ) = 0. In the limit of N → ∞, the global minimum will dominate the integrand and the integration range is largely determined by the region around x0 . Hence I ≈ lim e−N f (x0 )


N →∞


dx e 2 f


(x0 )(x−x0 )2


≈ lim e−N f (x0 ) N →∞

2π N f 00 (x0 )

1/2 . (1.91)

Task. Compute the next order correction to the saddle-point value of the integral. Coming back to our case, the result is given by the equation −


J −1 (x, y)φ0 (y) + tanh[βφ0 (x)] = 0 .



with the mean-field hφ(x)i = φ0 (x). A note on the mean-field approximation. At this point it is useful to make a comment on the notion of the mean-field approximation since the approach we took seems rather technical. In fact, there are several ways of 16 For

oscillating integrands, see the stationary phase approximation.


Computational Biology: A Statistical Mechanics Perspective

introducing a mean-field. A simple alternative is, e.g., to define and introduce b H(x) ≡


J(x, y)s(y)



as an effective field acting on the spins s(y). The resulting partition function then becomes one-dimensional and can be solved following the steps we took before. This procedure essentially means that within mean-field theory, one neglects correlations and factorizes hs(x)s(y)i = hs(x)ihs(y)i. While this procedure is technically much easier to perform than the computation we did, we will see in the following that our more systematic and general approach will also allow us, in a fairly straightforward sequence of steps, to learn something about the regime of validity of the mean-field approximation. We now return to eq.(1.92). If we assume that J(x, y) ≡ J, the system becomes translationally invariant and we can define the uniform order parameter M via φ0 = 2dJM (1.94) so that the saddle-point equation is rewritten as M = tanh(2dβJM ) .


The solutions of this equation can be obtained graphically, see Figure 1.1. If 2dβJM > 1, there is a pair of solutions which merge at M = 0 at the critical temperature βc−1 = 2dJ . (1.96) Within the mean-field approximation, the Ising model thus displays a phase transition at a finite temperature. This is at variance with our exact result for d = 1. So we have to clarify the limit of validity of the approximation we made. In fact, as we said before, the mean-field approximation assumes a factorization of correlation functions, hence a neglect of fluctuations.17 These can be accounted for by an expansion around the saddle-point value. We may thus speculate that the approximation is valid whenever this expansion is meaningful. In order to approach this question, we therefore have to pass on to a theory which is capable to capture the properties of the Ising model in the vicinity of the critical temperature. Ginzburg-Landau theory. In the vicinity of the transition, we can expand the terms in eq.(1.88) in a Taylor series (Why? Exercise!) 17 Why

do fluctuations destroy the factorization of correlation functions? Exercise.

Equilibrium Statistical Mechanics


FIGURE I.1.1: Phase transition in the Ising model in the mean-field approximation: graphical solution of the mean-field equation. A change in temperature allows to pass from a unique solution for T > Tc (bottom) to two solutions at T < Tc (top).

X x,y

J −1 φ(x)φ(y) = J −1




and ln[cosh(βφ(x))] =

 1 1 − 2 a2 ∇2 + ... φ(x) 2d 4d

β4 4 β2 2 φ (x) − φ (x) + ... 2 12



which, after passing to the continuum limit for a → 0 yields the GinzburgLandau form of the Hamiltonian   Z m2 2 λ 4 1 βH = dd x (∇φ)2 + φ + φ . (1.99) 2 2 4!


Computational Biology: A Statistical Mechanics Perspective

This theory can easily be treated. We consider the case of a homogeneous order parameter, φ = const. Then βV0 (φ) =

m2 2 λ φ + φ4 2 24


is the mean-field or effective potential of the Ginzburg-Landau theory. Depending on the sign of the quadratic term - with a fourth-order term which has to be strictly positive for thermodynamic stability - the potential displays two shapes which are shown in Figure 1.2. For m2 > 0, the potential has a single minimum at φ = 0; this can be identified with a disordered state. For m2 < 0, two minima appear at values φ = ± φ0 . They correspond to a pair of ordered states related to each other by a mirror-symmetry: one has a positive, the other a negative value of the same magnitude. For the ferromagnet these states correspond to states with positive and negative magnetization, and are identical to the solutions obtained from the graphical solution of the saddle-point equation in Figure 1.1. So far we have not specified the value of m2 ; in any case, in the vicinity of the transition at m = 0 we know that m2 ∼ βc − β ∼ |T − Tc | .


On the other hand, on dimensional grounds, m must be an inverse length, and we define m ≡ ξ −1 ∼ |T − Tc | 1/2 , (1.102) i.e., ξ ∼ |T − Tc | −1/2 .


We had encountered ξ before: it is the (spin-spin) correlation length. In eq.(1.103) we have obtained the first critical exponent which describes the power-law behaviour of physical quantities in the vicinity of the critical point: the exponent ν = 1/2 is thus the critical exponent of the correlation length. The information that the approach to a critical point is by a power-law is essential. Power laws are a general characteristic of systems without an intrinsic length scale (an example is a decay-length of correlations), or, in other words, in scale-invariant systems. This property of critical systems explains why power laws are so dearly loved by statistical physicists, and why they try to find them in more complex systems as well - a point we will return to in Part III of the book.

Equilibrium Statistical Mechanics


FIGURE I.1.2: Mean-field potential of the Ginzburg-Landau theory, for T < Tc (m2 < 0) and T > Tc (m2 > 0).

The Ginzburg-Landau model is, from a computational point of view, a very nice topic. We therefore suggest a number of exercises and tasks for the readers to try. Task. Perform the steps explicitly which are needed to pass from the discrete to the continuum version of the Ginzburg-Landau theory, leading to eq.(1.99). Verify that the mass parameter and the coupling constant are given by β −2 4β 2 m2 = c 2 (βc − β) , λ = 2 4 ad−4 . (1.104) Ja J βc

Exercise. Perform the mean-field approximation for a homogeneous order parameter at the level of the continuous Ginzburg-Landau theory by adding a field term −hφ. Discuss the phase transition in this model for h = 0 at T = Tc and at T < Tc and h → 0. As it turns out, the phase transition of the Ising model at T = Tc , h = 0 is a second-order phase transition, while for T < Tc , h → 0, it is a first-order phase transition. This terminology derives from the singular (power-law) behaviour of the first- and second derivatives of the free energy near the transition. At


Computational Biology: A Statistical Mechanics Perspective

a first-order phase transition, there is a jump in the first derivative of the free energy or, in other words, the free energy has a kink. For a second-order phase transition, the free energy is continuous and differentiable, but at the transition its second derivative displays a discontinuity. We will encounter the distinction between first- and second-order transitions as well in Part II, Chapter 2, in the discussion of the thermal stability of DNA. Exercise. Draw the effective potential of the Ising model in GinzburgLandau theory with and without field. Draw the shape of the potential near the first and second order transition to illustrate the above discussion. Exercise. Calculate the dependence of the susceptibility on the distance of the critical point, |T − Tc |, i.e. χ = ∂h φ0 |h=0 .


Deduce the critical exponent of the susceptibility. Exercise. Consider the phase transition from the disordered to the ordered state at h = 0 and at a finite value of h. Calculate the specific heat C ≡ −T ∂T F .


What do you notice? Determine the critical exponent. Task. Are there also solutions to the Ginzburg-Landau equations which depend on space, e.g., one-dimensional profiles φ(x)? What is their physical interpretation? Phenomenological scaling. Having found the first examples of critical exponents, the next obvious question to ask is whether there are any relations between them. In order to answer this question we start from the homogeneous Ginzburg-Landau equation with field and express it in the form m2 φ + uφ3 = h .


The solutions to this equation describe a family of solutions φ = φ(m2 , h). We have discussed before what happens if h = 0; if we approach the critical point at T = Tc , hence at m2 = 0, we find the power-law dependence φ ∼ h 1/3 ,


which yields another critical exponent. Is there a way to combine the two control parameters m2 and h? Let us introduce a scaled field variable ψ = φ h−1/3 .


Equilibrium Statistical Mechanics


Going back to eq.(1.107) we find m2 h−2/3 ψ + uψ 3 = 1 .


Thus by redefining x ≡ m2 h−2/3 we obtain xψ + uψ 3 = 1 and the whole dependence on m2 and h is now through a parameter combination. The general solution of the Ginzburg-Landau equation thus will have the scaling form φ(m2 , h) = h1/3 ψ(m2 h−2/3 ) . (1.111) This equation states that the function φh−1/3 has a universal form, and all curves parametrized by x will precisely collapse on this one universal curve. What is the expression of this function? For the cubic polynomial, one can calculate it explicitly (Task!), but in most cases this is not feasible. It is more instructive to see the properties of this function in certain limits. We know that at x = 0 ψ(0) ∼ u−1/3 . (1.112) For x → ∞, the cubic term in eq.(1.110) can be dropped and we find in that limit 1 (1.113) ψ(x) ≈ , x hence h h φ ≈ h1/3 (m2 h−2/3 )−1 = 2 ∼ (1.114) m T which is the high-temperature or weak-field limit; the final result is called Curie’s law of paramagnetism in the context of magnetism. For the opposite limit, x → −∞, we have |xψ|  1 and |uψ 3 |  1 so that p ψ ≈ |x|/u . (1.115) This we also have seen before, it is the ordered phase with φ ∼ |m2 | 1/2 . The idea to combine the control parameters into a single scaling variable has proved extremely fruitful for the theory of phase transitions. In the midsixties of the last century B. Widom proposed that quite generally the free energy per unit volume can be written in the scaling form f (m2 , h) = |m2 |2−α fs (h(m2 )∆ ) ,


where α is the specific heat exponent, and ∆ the gap exponent. From this hypothesis, a whole sequence of critical exponent relations can be deduced that


Computational Biology: A Statistical Mechanics Perspective

are now known to hold independent of spatial dimension (with some notable exceptions). Task. From the above considerations of phenomenological scaling for the Ginzburg-Landau theory, deduce the relation between the critical exponents β, α, and ∆. Here, β is not to be confused with the abbreviation of (kB T ) −1 ; it is the common notation for the exponent characterizing the temperature dependence of the order parameter via m ∼ |T − Tc |β . Beyond Ginzburg-Landau theory. Ginzburg-Landau theory has permitted us to very easily find the equilibrium states φ0 of the system by an almost trivial analytic calculation, the minimization of the effective potential V0 (φ), eq.(1.100). Further, we have found the critical exponents. But we still have no clue about the validity of this approach. In contrast to our calculation of the partition function of the one-dimensional Ising model, spatial dimensions do nowhere appear explicitly, and the basic conflict between the one-dimensional result (no phase transition at T > 0) and the GinzburgLandau result (a phase transition at T = Tc 6= 0) persists. In order to finally resolve that conflict we now have to look for the effects of fluctuations around the mean-field solution. We expect that they will modify the behaviour we have obtained. In order to compute the fluctuations around the mean-field solution we put φ = φ0 (x) + δφ and expand the Ginzburg-Landau Hamiltonian in δφ. This leads to a partition function Z = exp[−β(H0 + H1 )]


where βH0 is the GL-value, and βH1 is determined by a Gaussian integral over the δφ-fluctuations which can be carried out exactly. It reads   Z 1 2 1 d 2 2 d x ln −∇x + m + λφ0 (x) δ d (x − y)|x=y (1.118) βH1 = 2 2 With the Fourier representation of the δ−function18 we rewrite this as Z βH1 =

18 The

dd x


  dd k −ik·x 1 2 2 2 e ln −∇ + m + λφ (x) eik·x x (2π)d 2 0


Fourier representation of a function we use in the following is given by the integral Z dd k ik·x fb(x) = e f (k) . (1.119) (2π)d

Note that we frequently drop the b. and distinguish between a function and its Fourier transform by its argument only.

Equilibrium Statistical Mechanics where the term ∼ e−ik·y was first pulled through y = x. If φ0 = const. we find  Z dd k βH1 = ln k2 + m2 + (2π)d


to the left and then put to  1 2 λφ . 2 0


Eq.(1.121) is the main result of this paragraph, and its consequences will now be discussed. Obviously, this integral is not well-behaved for large k-values, indicating the breakdown of the continuum theory for small spatial scales.19 We can render the integral finite by introducing a cut-off Λ, but already here it becomes evident that spatial dimension now enters in the calculation. In d = 3 we obtain   3/2 λφ20 Λ 1 1 2 2 − (1.122) βH1 = m + λφ0 + O(Λα ) 4π 2 6π 2 where higher-order terms depending on Λ are summed up in the last term; further, terms which do not contain a dependence on the field have been dropped. The expression (1.122) can be understood as an additional contribution to the effective potential in mean-field which we recall had the form βV0 =

1 2 2 λ 4 m φ0 + φ0 . 2 4!


Expanding the root in βH1 up to fourth order in φ0 we obtain βH1 = −

m3 3 λφ20 Λ λmφ20 1 λ2 4 m + − − φ , 6π 4π 2 8π 64π m 0


where the first term can be ignored since it does not depend on φ0 . We can now restore the original form of the GL-potential by the redefinitions m2ef f = m2 − and λef f = λ + 19 This

λΛ λm + 2π 2 4π 3 λ2 . 8π m



phenomenon is called a UV-divergence in field theory since it occurs at large wavevectors, hence ‘high energies’. We do not follow this point here in all its consequences, since this goes far beyond what is attempted here. The limited ambition here is to illustrate that a continuum theory has to be considered carefully: it is potentially dangerous to extrapolate its results down to microscopic scales. This may not show up in a purely meanfield approach, but if one wants to go beyond to include fluctuation effects, surprises can happen.


Computational Biology: A Statistical Mechanics Perspective

We stress again that, although the effective potential in d = 3 has the same form as the effective potential in mean-field, the new result does depend both on spatial dimension (the calculation is only valid in d = 3) and on the cutoff Λ we used in calculating the integral. The vicinity of the transition: m2 → 0. We are now ready to take the final step and look at the effect of fluctuations at the critical point, or, more precisely, in the vicinity of the phase transition. In order to do this we have to look at the correlation function of the field φ(x) in the limit m2 → 0, since this will now give us additional information beyond the value of the order parameter itself, which we can compute from the fluctuation-modified effective potential. Calling the correlation function of the field20 G(x), it fulfills the differential equation (−∇2x + m2 )G(x) = δ d (x) , (1.127) and, consequently, its Fourier transform is given by G(k) =


1 . + m2


with, as before, m ≡ ξ −1 ∼ |T − Tc | 1/2 . Now let’s transform back to real space. We find Z G(r) =

dd k eik·x = ξ 2−d (2π)d k2 + ξ −2


dd q eiq·x/ξ , (2π)d q2 + 1

where q = ξk. The integrand has two asymptotic limits for |x| = r:  −r/ξ e  r(d−1)/2 , rξ G(r) ∼  1 , r  ξ. r d−2



Note that the first expression indeed reduces to our previous result, eq.(1.75), for d = 1. This means that, within Ginzburg-Landau theory, although there is a mean-field transition, fluctuations will destroy this transition and the original (exact) calculation is supported.

20 We

made use of translational invariance of the system by shifting one of the arguments to x = 0. In a translationally invariant system, the correlation function depends on the distance of field at the two selected points in space - we saw this already before in our computation of the correlation function in the one-dimensional spin chain.

Equilibrium Statistical Mechanics


The existence of a second regime in which correlations decay algebraically signals the presence of a true ordered phase at a finite temperature, i.e., for 0 < T < Tc , and we see that this is certainly possible for d > 2. With this information we can now, finally, estimate the range of validity of the Ginzburg-Landau theory. Considering length scales on the order of the correlation length, we have G(ξ) ∼ ξ 2−d ∼ |T − Tc |

d−2 2



If we look at the ratio of G and the square of the mean-field value of the order parameter, d−4 G(ξ) (1.132) ∼ |T − Tc | 2 , φ20 we see that, in the vicinity of the transition for T → Tc , fluctuations grow indefinitely for d < 4. The dimension du = 4 is hence considered as an upper critical dimension above which mean-field theory becomes exact with respect to the critical exponents, i.e., fluctuations are negligible.21 The dimension dl = 1 is likewise a lower critical dimension at which the phase transition is destroyed by fluctuations. This concludes our discussion of the Ising model. We leave this model at a point when, from the point of view of statistical physics, things become really interesting: how can we mathematically describe the transition in the range of dimensions 1 < d < 4? In d = 2, the Ising model can be solved exactly by a transfer matrix approach (L. Onsager, 1944); in d = 3 the problem of the computation of the partition function was shown to be NP-complete (S. Istrail, 2000). It is here where renormalization-group methods need to be used, which go beyond what we attempt here. For those who are interested to get a systematic view of the theory of critical phenomena, some corresponding literature is listed at the end of this Chapter. Before moving on to our second exemplary system, the model of an elastic polymer, it is nevertheless important to put the calculations we went through into context. What we did looked very much like hard-core statistical physics, and we seem to have lost track of biology. So what is this all good for, in the context of computational biology? The answer that can be given to this question has three aspects. First, as we will see in this book, there are phase transitions in biological systems. The methodology we have developed here is thus needed to understand these 21 In

a more detailed calculation one can see that right at the upper critical dimensions, logarithmic corrections arise.


Computational Biology: A Statistical Mechanics Perspective

transitions. We will see a particular example in Part II of the book when we study the thermal denaturation of DNA. Even before, in Chapter 2 of Part I, we will see that the concept of a phase transition carries over also to nonequilibrium situations. The alignment of DNA or protein sequences can be understood as a nonequilibrium phase transition. Second, we will see in what follows that even though biological systems do not always have phase transitions, they may nevertheless exhibit scaling behaviour in some of their characteristic quantities. Scaling is thus a useful method of analysis, although one should not be carried away to overinterpret its outcome. We will touch on this aspect in Part II of the book, e.g., in the context of the analysis of microarray data, and in Part III when we discuss global features of biological networks. And finally, a phase transition is the prime example of a cooperative phenomenon. Cooperative phenomena arise in biology in various different contexts, and we will encounter them here, too. We will see a direct biological application of the Ising model to the cooperative behaviour of receptors in the membrane of a bacterium. We are now ready to turn to our second illustrative example, a model for an elastic (bio-)polymer. Modelling biopolymer elasticity. As a second application example for the methods of statistical mechanics we introduce a model description for the elastic behaviour of biomolecules. Molecular elasticity is entropic in origin. This merits an explanation. Let’s begin with a simplistic version of a polymer chain model which is called the freely jointed chain. We imagine the polymer can be abstracted as a chain of N freely jointed links, as illustrated in Figure 1.3. The spatial configuration of the chain resembles the trajectory of a random walk. If we now pull at the ends of our caricature polymer with a force F, we will straighten out the chain. What happens if we then let go? Imagining we have realized the chain as made from paperclips, it would simply stay straight, as there is no internal or external mechanism that would pull the segment back into the initial random configuration. Something is missing, this idea is wrong. A biomolecule in solution is subjected to thermal energy, kB T . The positional fluctuations of the surrounding molecules hit the chain and will thereby gently randomize the configuration. Straightening out the chain against these thermally fluctuating forces consequently reduces the entropy of the chain, i.e., the number of conformations it can adopt. Work has to be performed, and the relation between force and extension of a molecule in solution results

Equilibrium Statistical Mechanics


FIGURE I.1.3: The simplistic polymer model: a chain of linear segments which can freely move about their links.

thus from an entropic elasticity. Pulling DNA. Based on this concept, various experimental techniques have been developed to exert forces on molecules in solution. The basic principle is that the molecule has to be fixed at one end, while the other can be manipulated by mechanical force. A typical setup is that a DNA molecule is chemically modified at its both extremities; it is then rigidly fixed at one end to a surface (a capillary), and to a superparamagnetic bead at the other end. The application of a magnetic field allows to pull on the bead and hence the DNA molecule. Another possible setup is by trapping a bead in an optical trap; such a setup is indicated schematically in Figure 1.4. Questions of biological or biophysical interest that have been addressed with this approach are listed in the additional notes at the end of this chapter; here we only want to understand how to describe the elastic properties. The WLC-model. Following the suggested discrete representation of the polymer as a chain of elements, we first consider the polymer as a linear chain with element length b, such that the chain made of N segments has a total length of L = bN . The energy of the chain is given by the Hamiltonian βH = −K

X i

b ti · b ti+1 − βF b


b ti · zb



where an applied force F is directed along the z-axis. In eq.(1.133), K/β is the stiffness of the chain, i.e., its resistance to bending. The unit vector b ti describes the orientation of the segment i. This model is sometimes called the Kratky-Porod model, which for K = 0 reduces to the


Computational Biology: A Statistical Mechanics Perspective

FIGURE I.1.4: trap.

A setup for DNA stretching experiments using an optical

freely jointed chain we had pictured before. In the continuum limit b → 0, where K ≡ `P /b introduces the persistence length `P , we can represent the model as the Worm-Like Chain (WLC) model with the energy expression βHW LC =

`P 2

Z 0


ds(∂sb t )2 − f

Z 0


ds b t · zb


where the pulling force per unit length is F = kB T f . In this formulation, s is the curvilinear coordinate along the chain, and b t has turned from the orientational vector of a chain element into the tangential vector along the continuous curve which makes out the chain. As noted, the first term in the Hamiltonian is the bending energy. Its coefficient, the persistence length, is a length scale measure for the correlations of the tangent vectors along the chain, or, in other words, for the propagation of changes in the chain conformation along the chain.22

22 There

are several other commonly used lengths that characterize polymers; a prominent one is the Kuhn length with `K = 2`p , which is well-defined provided L  `K .

Equilibrium Statistical Mechanics


As stated above, the quantity we are interested in from an experimental point of view is the force-extension relation. For the WLC-model, the exact computation of the partition function is equivalent to solving the problem of a quantum rotator subjected to a polarizing field. Here we give a simplified calculation valid for the case when the forces pulling at the molecule are sufficiently strong and limit the chain excursions transverse to the pulling direction (J. F. Marko and E. Siggia, 1995). Since |b t| = 1, we can take the transverse components (tx , ty ) ≡ b t⊥ as independent variables, and obtain for the z-component tz = 1 − b t2⊥ /2 + O(b t4⊥ ) .


We can then express the energy in a Gaussian approximation as βHW LC =

1 2



ds[`p (∂sb t⊥ )2 + f b t2⊥ ] − f L .



Introducing the Fourier modes ˜ t⊥ (q) = we obtain βHW LC =

1 2



ds eiqs b t⊥ (s)

dq [`p q 2 + f ]˜ t2⊥ − f L . 2π



The average of the transverse components along the chain s follows from equipartition   Z Z dq 1 1 dq ˜ 2 2 b h|t⊥ (q)| i = 2 =√ (1.139) ht⊥ i = 2π 2π `P q 2 + f f `P t⊥ . The extension where the factor of two accounts for the two components of b of the chain is then obtained as √ z hb t2⊥ i f `P b = t·b z=1− =1− (1.140) L 2 2 which behaves as a square-root in f . The approximate formula   1 z 1 1 f= + − `P L 4(1 − z/L)2 4 describes the experimental behaviour rather well, see Figure 1.5.



Computational Biology: A Statistical Mechanics Perspective

FIGURE I.1.5: Force-extension curve for a single-molecule DNA. (Reprinted from C. Bustamante et al., 2000, with permission from Elsevier.)

Equilibrium Statistical Mechanics


Additional Notes Equilibrium statistical mechanics is a well-developed theory on which many books have been and are being published, written to every taste of mathematical rigor. Four suggestions are the books by K. Huang, M. Plischke and B. Birgersen, F. Schwabl and L. E. Reichl. Books focussing on the renormalization group methods of statistical physics are, again as examples, the volumes by L. P. Kadanoff and N. Goldenfeld. A highly recommendable introduction into biophysics, touching on many aspects of statistical physics, has been written by P. Nelson. Experiments on single molecule manipulations of nucleic acids have started in the early nineties of the last century; basic references on DNA elasticity are T. Strick et al, 1996 and 1998. Reviews on the basic physics and the experimental techniques are C. Bustamante et al., 2000, and U. Bockelmann, 2004. The advantage of the single-molecule technique is that the mechanical force exerted on the molecule can be made to interfere with enzymatic reactions on the chains, allowing for a mechanical probing of biochemical processes. Examples from the catalogue of achievements are: • Unwinding of promoter regions in transcription initiation: A. Revjakin et al., 2004. • Pausing of RNA polymerase during transcription: R. J. Davenport et al., 2000 and N. R. Forde et al., 2002. • Base-pair stepping of RNA polymerase: E. A. Abbondanzieri et al., 2005. • Torque-induced structural transitions in DNA, of relevance for enzymatic reactions: Z. Bryant et al., 2003. • DNA unzipping: C. Danilowicz et al., 2003. Theoretical papers on DNA elasticity are the classic paper on the wormlike chain model by J. F. Marko and E. D. Siggia, 1995. Extensions to ´zard, 1997, and include twist were published by C. Bouchiat and M. Me J. D. Moroz and P. Nelson, 1997.


Computational Biology: A Statistical Mechanics Perspective References

E. A. Abbondanzieri, W. J. Greenleaf, J. W. Shaevitz, R. Landick and S. M. Block, Direct observation of base-pair stepping by RNA polymerase, Nature 438, 460-465 (2005) U. Bockelmann, Single-molecule manipulation of nucleic acids, Curr. Op. Struct. Biol. 14, 368-373 (2004) C. Bouchiat and M. M´ezard, Elasticity Model of a Supercoiled DNA Molecule, Phys. Rev. Lett. 80, 1556-1559 (1998) Z. Bryant, M. D. Stone, J. Gore, S. B. Smith, N. R. Cozzarelli and C. Bustamante, Structural transitions and elasticity from torque measurements on DNA, Nature 424, 338-341 (2003) C. Bustamante, S. B. Smith, J. Liphardt and D. Smith, Single-molecule studies of DNA mechanics, Curr. Op. Struct. Biol. 10, 279-285 (2000) C. Danilowicz, V. W. Coljee, C. Bouzigues, D. K. Lubensky, D. R. Nelson and M. Prentiss, DNA unzipped under a constant force exhibits multiple metastable intermediates, Proc. Natl. Acad. Sci. USA 100, 1694-1699 (2003) R. J. Davenport, G. J. L. Wuite, R. Landick and C. Bustamante, SingleMolecule Study of Transcriptional Pausing and Arrest by E. coli RNA Polymerase, Science 287, 2497-2500 (2000) N. R. Forde, D. Izhaky, G. R. Woodcock, G. J. L. Wuite and C. Bustamante, Using mechanical force to probe the mechanism of pausing and arrest during continuous elongation by Escherichia coli RNA polymerase, Proc. Natl. Acad. Sci. USA 99, 11682-11687 (2002) N. Goldenfeld, Lectures on Phase Transitions and the Renormalization Group, Westview Press (1992) K. Huang, Introduction to Statistical Physics, CRC Press (2001) S. Istrail, Statistical Mechanics, Three-Dimensionality and NP-Completeness: I. Universality of Intractability of the Partition Functions of the Ising Model Across Non-Planar Lattices, Proc. 32nd ACM Symposium on the Theory of Computing (STOC00), ACM Press, 87-96 (2000) E. T. Jaynes, Information Theory and Statistical Mechanics I., Phys. Rev. 106, 620-630 (1957)

Equilibrium Statistical Mechanics


E. T. Jaynes, Information Theory and Statistical Mechanics II., Phys. Rev. 108, 171-190 (1957) L. P. Kadanoff, Statistical Physics. Statics, Dynamics and Renormalization, World Scientific (2000) J. Majewski, H. Li and J. Ott, The Ising Model in Physics and Statistical Genetics, Am. J. Hum. Genet. 69, 853-862 (2001) J. F. Marko and E. D. Siggia, Stretching DNA, Macromolecules 28, 87598770 (1995) J. D. Moroz and P. Nelson, Torsional directed walks, entropic elasticity, and DNA twist stiffness, Proc. Natl. Acad. Sci. USA 94, 14418-14422 (1997) P. Nelson, Biological Physics. Energy, Information, Life, Freeman (2004) L. Onsager, Crystal statistics. I. A two-dimensional model with an orderdisorder transition, Phys. Rev. 65, 117-149 (1944) M. Plischke and B. Birgersen, Equilibrium Statistical Physics, Scientific Publishers 2nd ed. (1994) L. E. Reichl, A Modern Course in Statistical Physics, Wiley-Interscience 2nd ed. (1998) A. Revjakin, R. H. Ebright and T. R. Strick, Promoter unwinding and promoter clearance by RNA polymerase: Detection by single-molecule DNA nanomanipulation, Proc. Natl. Acad. Sci. USA 101, 4776-4780 (2004) F. Schwabl, Statistical Mechanics, Springer (2002) T. R. Strick, J.-F. Allemand, D. Bensimon, A. Bensimon and V. Croquette, The Elasticity of a Single Supercoiled DNA Molecule, Science 271, 1835-1837 (1996) T. R. Strick, J.-F. Allemand, V. Croquette and D. Bensimon, Physical Approaches to the Study of DNA, J. Stat. Phys. 93, 647-672 (1998) B. Widom, Equation of State in the Neighborhood of the Critical Point, J. Chem. Phys. 43, 3898-3905 (1965)

Chapter 2 Nonequilibrium Statistical Mechanics

Nonequilibrium statistical mechanics does not yet have such powerful general concepts and tools as equilibrium statistical mechanics. Attempts to develop such concepts for physical processes which occur arbitrarily far from a thermal equilibrium situation have so far not led to the desired success.1 Since we still lack these general principles which govern the time-dependent or stationary distributions of our physical quantities of interest, we instead have to directly address the properties of the stochastic processes themselves and try to find methods to treat them, at least with some reasonable simplifications. In the most general case we can address, we will be able to compute the time evolution of suitably defined probability distributions characterizing the state of a physical system, or its transition from one state to another. Consequently, our symbol of choice in this Part is P rather than Z. What we will also see, however, is that we sometimes will be able to relate nonequilibrium processes to equilibrium processes. We will find expressions, the so-called (dissipation-)fluctuation theorems, which establish such a relation. Most often, however, these relations require for their validity to be not too far away from a thermal equilibrium. There is one notable exception we will address as well, a result obtained some years ago by C. Jarzynski. We begin this Chapter with the formal bits and pieces of a general description of stochastic processes and their time evolution.2

1 Of

course, there is an extended body of work on irreversible processes. Nevertheless, I think there is a general consensus that a theory of noneqilibrium statistical physics of the same level of generality as that of equilibrium statistical physics has not yet been established. 2 The presentation, in particular at the beginning, follows the book by N. van Kampen, 1992, which can be seen as a standard reference on most of the topics discussed in this Chapter.



Computational Biology: A Statistical Mechanics Perspective


Stochastic processes

The first section of this Chapter lists a number of definitions for later use. We call X be a stochastic variable with a particular value x; f is considered a mapping from X at time t. We call YX (t) = f (X, t)


a random function, and Yx (t) = f (x, t) a sample function or a realization of the stochastic process; as for equilibrium states, we can speak of the corresponding ensemble in a natural way. For the stochastic process, we define an ensemble average by Z hY (t)i ≡ dx Yx (t) P (x) . (2.2) The higher moments of the distribution are defined in an analogous way. The probability for Yx (t) to take the value y at time t is given by Z P (y, t) = dxP (x)δ(y − Yx (t)) . (2.3) The mean and the moments are, as in thermal equilibrium, quantities that allow to quantify the probability distributions. For time-dependent processes, another quantity is of interest: The autocorrelation function. The autocorrelation function of the stochastic process is defined by A(t1 , t2 ) ≡ hY (t1 )Y (t2 )i − hY (t1 )ihY (t2 )i .


For t1 = t2 = t, it reduces to the time-dependent variance, σ 2 (t). Joint probability density. We define the joint probability density P (y1 , t1 ; ...; yn ; tn )


which states that Yx (t) has the value y1 at t1 , ..., yn at tn . Exercise. Write down the mathematical expression for P . Conditional probability. The conditional probability is the probability density for Yx (t) to take on the value y2 at t2 if it was y1 at t1 , with the normalized density Z dy2 P (y2 , t2 |y1 , t1 ) = 1 .


Nonequilibrium Statistical Mechanics


Markov process. A stochastic process has the Markov property if for any set of ordered timesteps t1 < t2 < ... < tn the conditional probability satisfies P (yn , tn |y1 , t1 , ..., yn−1 , tn−1 ) = P (yn , tn |yn−1 , tn−1 ) .


P is then a transition probability. It only depends on the two states involved, the one that is left and the one that is reached. A Markov process is uniquely determined from the knowledge of P (y1 , t1 ) and P (y2 , t2 |y1 , t1 ). Exercise. Convince yourself of the correctness of the last statement for a process involving three steps, t1 < t2 < t3 . The Chapman-Kolmogorov equation. From the transition probabilities of a Markov process we have the expression Z P (y3 , t3 |y1 , t1 ) = dy2 P (y3 , t3 |y2 , t2 )P (y2 , t2 |y1 , t1 ) (2.8) in which time-ordering is essential. Examples. a) For −∞ < y < ∞, the Chapman-Kolmogorov equation is solved by (for t2 > t1 )   1 (y2 − y1 )2 P (y2 , t2 |y1 , t1 ) = p exp − (2.9) 2(t2 − t1 ) 2π(t2 − t1 ) If P (y, 0) = δ(y), this Markov process is called the Wiener or Wiener-L´evy process. b) If Yx (t) takes on only positive integer values n = 0, 1, 2, ... for t ≥ 0, eq.(2.8) is obeyed by the Poisson process n −n1

P (n2 , t2 |n1 , t1 ) =

(t2 − t1 ) 2 (n2 − n1 )!

e −(t2 −t1 )


and P (n, 0) = δn,0 .


Task. Show that the probability density of the Wiener process fulfills the diffusion equation ∂t P (x, t) = D ∂x2 P (x, t) (2.12) with a diffusion constant D = 1/2.


Computational Biology: A Statistical Mechanics Perspective

Stationary Markov processes. A stochastic process is called stationary if all joint probability densities depend only on time-differences τ P (y1 , t1 ; ....; yn , tn ) = P (y1 , t1 + τ ; ...; yn , tn + τ ) .


In this case we write for the conditional probabilities P (y2 , t2 |y1 , t1 ) = Tτ (y2 |y1 )


and we will also suppress the index τ whenever no ambiguity can arise. With this notation, the Chapman-Kolmogorov equation is rewritten as Z T (y3 |y1 ) = dy2 T (y3 |y2 )T (y2 |y1 ) . (2.15)

Task. Show that the autocorrelation function of a stationary Markov process with zero mean is given by Z Z A(τ ) = dy1 dy2 y1 y2 Tτ (y2 |y1 ) P (y1 ) (2.16) for τ ≥ 0. Example. The standard example of a stationary Markov process is the Ornstein-Uhlenbeck process, for which   1 y2 P (y1 ) = √ exp − 1 (2.17) 2 2π and

  (y2 − y1 e−τ )2 exp − 2(1 − e −2τ ) 2π(1 − e−2τ )

T (y2 |y1 ) = p



where τ ≡ t2 − t1 . Exercise. Show that the Ornstein-Uhlenbeck process fulfills hY i = 0 and A(τ ) = e −τ .


The master equation

With this minimal list of definitions we have set the stage for the description of stochastic processes within the context of statistical physics, and are

Nonequilibrium Statistical Mechanics


ready to move on.3 The next and very important step is to make practical use of the Chapman-Kolmogorov equation for more complex situations than we have discussed so far in the few examples. The corresponding mathematical tool is the master equation. The master equation is a limiting expression one can obtain from the Chapman-Kolmogorov equation if one lets the time interval tend to zero in a controlled way such that one can pass over to a differential equation for the transition probabilities. Let’s do this. Suppose we write down the transition probability T for the case of small time intervals τ 0 in the following way 2

Tτ 0 (y2 |y1 ) = (1 − a0 (y1 )τ 0 )δ(y2 − y1 ) + τ 0 W (y2 |y1 ) + o(τ 0 ) .


This is an expansion to linear order in τ 0 , in which W (y2 |y1 ) is defined as the transition probability per unit time from y1 to y2 , hence W (y2 |y1 ) ≥ 0. The second term in eq.(2.19) is thus clear, but we have still to say something about the first term. In this term, the coefficient in front of the δ-function takes into account that no transition occurs during the time interval τ 0 , and hence we must have for the coefficient a0 Z a0 (y1 ) = dy2 W (y2 |y1 ) . (2.20) Finally, the last term in eq.(2.19) denotes terms of higher order in τ 0 which we neglect in the following. We insert this expression into the Chapman-Kolmogorov equation and do some rearrangement of terms, with the result Z Tτ +τ 0 (y3 |y1 ) − Tτ (y3 |y1 ) = −a (y ) T (y |y ) + dy2 W (y3 |y2 )Tτ (y2 |y1 ) 0 3 τ 3 1 τ0 (2.21) which in the limit τ 0 → 0 yields the differential equation Z ∂τ Tτ (y3 |y1 ) = dy2 [W (y3 |y2 )Tτ (y2 |y1 ) − W (y2 |y3 )Tτ (y3 |y1 )] .


This is the master equation in its continuous form. Usually it is rewritten in 3A

reader who wishes to get more is asked to consult van Kampen’s book; see the list of references.


Computational Biology: A Statistical Mechanics Perspective

a more intuitive form as Z ∂t P (y, t) = dy 0 [W (y|y 0 )P (y 0 , t) − W (y 0 |y)P (y, t)] .


This equation must be solved for t ≥ t1 given an initial condition P (y, t1 ) = δ(t − t1 ). Note that the equation should not be misinterpreted as an equation for a single-time distribution, which is a frequent error due to the abusive notation with P ; the Chapman-Kolmogorov equation makes it clear that we are considering transition probabilities. Having our tool finally at hand, we can now turn to some illustrative examples. For this, we employ the master equation in a form applicable to discrete states n, and write it as X p˙n (t) = [wnm pm (t) − wmn pn (t)] . (2.24) m

This equation can also be understood as a gain-loss equation for the probability of states n with wnm ≥ 0. Example. As an example of a non-stationary Markov process, we consider protein degradation. Within a cell, proteins are continually degraded by a dedicated cell machinery. In our simple model we describe this by a rate γ. The transition probability is given by wnm = γ m δn,m−1


with n 6= m. The master equation of this process reads p˙n = γ(n + 1)pn+1 (t) − γnpn (t)


which needs to be solved under the initial condition pn (0) = δn,n0 , i.e., we assume that at time t = 0 there are n0 proteins present. The simple way to ‘solve’ eq.(2.26) is to consider the evolution of the average number of proteins, hn(t)i. This is done by multiplying the equation by n and summing up ∞ X

np˙n = γ


∞ X n=0 ∞ X

n(n + 1)pn+1 − γ

∞ X

n 2 pn


(n − 1)npn − γ

n=0 ∞ X

= −γ


∞ X n=0


n 2 pn


Nonequilibrium Statistical Mechanics


which is nothing but d hn(t)i = −γhn(t)i dt


hn(t)i = n0 e −γt .


solved by

This is, of course, not the full solution to the problem - but it nicely shows how the behaviour of the average can be obtained. In order to really solve the master equation we can make use of the generating function n0 X G(s, t) = pn (t)sn (2.30) n=0

defined for |s| ≤ 1. Multiplying eq.(2.26) by sn and summing up, the master equation is transformed into a first-order partial differential equation for G in the variables s, t, ∂t G(s, t) + γ(s − 1)∂s G(s, t) = 0 .


The form of the equation indicates that a separation of variables will be of help. We are thus led to write the following ansatz4 for G G(s, t) = (a(t)(s − 1) + b)n0 ,


where the function a(t) and the constant b need to be determined. Plugging the ansatz into eq.(2.31), the s-dependence drops out, and the function a(t) is found to fulfill the simple differential equation (compare to eq.(2.28)) a(t) ˙ = −γ a(t)


with the solution a(t) = e−γt . The constant b = 1, as follows from the initial condition G(s, 0) = sn0 . Thus G(s, t) = (1 + e−γt (s − 1))n0 .

4 Note


that according to its definition, G is a polynomial in the variable s of order n0 .


Computational Biology: A Statistical Mechanics Perspective

Now we have two expressions for G(s, t), the power series in s and its sum. Since the pn (t) are the coefficients of the series, we only have to Taylor expand eq.(2.34), which leads to the final result   n0 pn (t) = exp(−γnt)(1 − exp(−γt))n0 −n . (2.35) n

One-step processes. A frequently occurring class of stochastic processes that can be studied with the help of master equations are one-step processes in which transitions occur only between neighbouring state labels. Writing the master equation as a matrix equation5 p˙n = Wnm pm


Wnm = fm δn,m−1 + bm δn,m+1


we define the matrix Wnm as

with n 6= m. The diagonal element of the matrix is given by Wnn = −(fn + bn )


so that the full equation reads as p˙n = fn+1 pn+1 + bn−1 pn−1 − (fn + bn )pn .


The last equation can be written in a more concise form using the step operator E, defined by its action to “move up or down the ladder” of states given by hn : Ehn = hn+1 , E −1 hn = hn−1 (2.40) so that for eq.(2.39) we have p˙n = (E − 1)fn pn + (E −1 − 1)bn pn .


Exercise. Solve the master equation (2.39) for the symmetric random walk with fn = bn = 1, p˙n = pn+1 − pn−1 − 2pn , −∞ < n < ∞ for the initial data pn (0) = δn,0 .

5 Which

makes the linearity of the equation evident.


Nonequilibrium Statistical Mechanics



Fluctuation theorems

We have mentioned in the introduction to this Chapter that nonequilibrium statistical mechanics still lacks general concepts when compared to equilibrium statistical mechanics. In this section we will derive a theorem (or, an expression) which describes a key property of nonequilibrium systems. The fluctuation-dissipation theorem. We derive a general fluctuation theorem, valid for nonequilibrium stationary states (and, a fortiori, near equilibrium states) within the formalism we have introduced (M. Lax, 1960). Suppose we have I molecular species present with numbers ni , 1 ≤ i ≤ I; we consider the index i as a vector index on particle numbers, and abbreviate n = (n1 , ...nI ) = ni , which is hence a row vector. From our previous results we infer that the particle distribution function for a Markov process fulfills Z P (n, t + ∆t) = dmP (n, t + ∆t|m, t)P (m|t) , (2.43) We now define the following quantities6 Z 1 A(m) = dnP (n, t + ∆t|m, t)(n − m) , ∆t


which is a vector; we call it a drift vector.7 Further, we define the matrix Z 1 dnP (n, t + ∆t|m, t)(n − m) · (n − m)T (2.45) D(m) = 2∆t where (...)T denotes the transpose of the row vector of particle numbers, hence a column-vector. We call D a diffusion matrix. Supposing that the stationary state is characterized by a particle number vector n0 , we introduce the vector δn ≡ n − n0 .


We now find (verify this as a Task) the equation for the time evolution of the mean-value d hδni = hA(n(t))i (2.47) dt 6 This 7 The

step assumes that the conditional probabilities are expandable to linear order in ∆t. names of the objects will become clear later.


Computational Biology: A Statistical Mechanics Perspective

and the corresponding equation for the covariance matrix d hδn δnT i = 2hD(n)i + hA(n)δnT i + hδnAT (n)i . dt


As a following step we expand around the stationary state via n = n0 + δn and assume for the drift-vector A(n) ≈ A(n0 ) − Λδn ,


where we have introduced a matrix Λ. In this last equation we choose A(n0 ) = 0


to be consistent with eq.(2.47). Further, we make the simplifying assumption that D(n) ≈ D(n0 ) ≡ D . (2.51) We then end up with the following results for the mean d hδni = −Λhδni dt


d hδnδnT i = 2D − ΛhδnδnT i − hδnδnT iΛT . dt


and the covariance matrix

These equations are the main result of this section. We note that in steadystate, a relation between the diffusion matrix D, the fluctuations in particle number δn and the drift matrix Λ is established in the form 2D = ΛhδnδnT i + hδnδnT iΛT .


This equation is a general version of the so-called dissipation-fluctuation theorem. The name derives from special cases - to which we will come in the following - in which the drift vector corresponds, e.g., to a friction force acting on a particle and hence provides a mechanism for dissipation. The Jarzynski equality. We now turn to another general result, the socalled Jarzynski equality. This time we start the derivation from the master equation, following U. Seifert, 2004. We consider a situation in which we allow the wmn in the master equation to depend on some tunable parameter λ, i.e., we have wmn = wmn (λ) .


Nonequilibrium Statistical Mechanics


For a fixed value of λ, we assume (as before) that the system is in a stationary state psn which obeys the condition of detailed balance. This condition is quite important: while obeyed by equilibrium systems, the reverse it not true. The condition of detailed balance suffices for the system to have a stationary state, and hence is an important information to have on a nonequilibrium system. Mathematically, the condition of detailed balance is given by wmn psn = . s pm wnm


If we now assume that the parameter λ is turned on in a time-dependent manner, λ = λ(τ ), we would like to know the probability P to encounter a particular trajectory of the system n(τ ) ≡ (n0 , n1 , ..., nk ) starting in n0 at time τ0 = 0, jumping to n1 after a time-interval τ1 and so forth, until the final jump from τk−1 to τk ≡ t. This probability is given by P [n(τ ), λ(τ )] = ps (n0 , λ(0)) ×

k−1 Y

 Z exp −

dτ τi




wn0 ,m (λ(τ ))


×wn0 ,m (λτi+1 ) .


Likewise, we can study the trajectory n ˜ ≡ n(t−τ ) which occurs under reversal ˜ This operation allows to write down the probability of λ, i.e., λ(t − τ ) ≡ λ. ˜ )] (Exercise). P [˜ n(τ ), λ(τ We can then form the ratio of the two probabilities which is given by e


 Z t  ˜ )] P [˜ n(τ ), λ(τ 0 ˙ = exp − dτ  n(τ ) λ(τ ) ≡ P [n(τ ), λ(τ )] 0


where the quantity 0 n (λ) = −

d ln ps (n, λ) dλ


has been introduced. It can be considered as a formal ‘energy level’. This interpretation can be made if one wishes to read the integral in the argument of the exponential function as an analogue of a free energy - we will soon see that this analogy can be made precise. From these observations we obtain the following two identities 1=

X n ˜ (τ )

˜ )] = P [˜ n(τ ), λ(τ

X n ˜ (τ )

e −R[n(τ )] P [n(τ ), λ(τ )]



Computational Biology: A Statistical Mechanics Perspective

and 1=



−R[n(τ )]

 P [n(τ ), λ(τ )] = hexp

Z −


 ˙ dτ  n(τ ) λ(τ ) i . 0



n(τ )

We interpret the second expression as a fluctuation theorem  Z t  ˙ ) i = 1, hexp − dτ 0 n(τ ) λ(τ



and explain now why, by some illustrative applications. Illustration of the fluctuation theorem. We first illustrate the result eq.(2.62) by applying it to a simple cyclically working enzyme or motor, following U. Seifert, 2005, as depicted in Figure 2.1. The enzyme is assumed to have three equivalent conformational states, and it progresses from one to the other at a rate k+ in the forward (i.e., clockwise), and with a rate k− in the backward direction. We assume k+ > k− . The stationary distribution of the system clearly is given by ps = 1/3: the system spends equal times in each of the states. For this system we obtain R as R = n ln(k+ /k− )


where n ≡ n+ − n− is the effective number of steps in the forward direction. Thus  n k− P [−n] −n ln(k+ /k− ) =e = . (2.64) P [n] k+ The exact P [n] can be computed from the master equation, since this system is an asymmetric random walk with the master equation p˙n = k+ pn+1 − k− pn−1 − (k+ + k− )pn


for which p pn ≡ P [n] = I|n| (2 k+ k− t)

k+ k−


e −(k+ +k− )t ,


where In (x) is the modified Bessel function of order8 n. One sees that the factor (k− /k+ )n arises from the ratio of backwards and forward processes with the probability ratio P [−n]/P [n].

8 The

modified Bessel function of order n is given by the expression Iα (x) = i−α Jα (x)

with Jα (x) =

∞ X m=0

 x 2m+α (−1)m . m!Γ(m + α + 1) 2

Nonequilibrium Statistical Mechanics


FIGURE I.2.1: An enzyme or molecular motor switching between three configurational states. Transition rates in forward and backward direction are given by ki+ and ki− for i = 1, 2, 3, respectively (after U. Seifert, 2005). Jarzynszki theorem. The fluctuation theorem above can further be elucidated by pointing out its relation of eq.(2.62) with a theorem due to C. Jarzynski, 1997, as suggested by U. Seifert, 2004. Consider the trajectory of a particle x(t) in a potential V (x, λ), which again depends on the parameter λ. The stationary distribution of this process for fixed λ is given by ps (x, λ) = Z(λ) −1 exp[−βV (x, λ)]


with the normalization given by the ‘partition function’ Z ∞ Z(λ) ≡ dx exp[−V (x, λ)] .



Substituting these correspondences into the formula (2.59) one has with n(t) ∼ x(t) n (λ) ∼ − ln ps (x, λ) = βV (x, λ) + ln Z(λ) . (2.69) Inserted into eq.(2.62) this leads to the result


Computational Biology: A Statistical Mechanics Perspective   Z t ˙ )) Z(λ(0)) i = 1 . hexp −β dτ V 0 (x(τ ), λ(τ Z(λ(t)) 0


Invoking the expression for the free energy F (λ) = β −1 ln Z(λ) , and introducing the notions of the applied work Z t ˙ ) W ≡ dτ V 0 [x(τ ), λ(τ )]λ(τ




and the dissipated work Wdiss ≡ W − [F (λ(t)) − F (λ(0))]


we finally arrive at the expression he −βWd i = 1


he −βW i = e−β∆F


or, equivalently, at

where ∆F can be read off from eq.(2.73). This is Jarzynski’s original result (1997). The Jarzynski equation (2.75) is remarkable since it relates the difference between an equilibrium free-energy difference to the average over a nonequilibrium quantity. This merits a deeper discussion. Validation of Jarzynski’s equality. Jarzynski’s result can be rewritten as an equation for the free energy difference in terms of a nonequilibrium expression ∆F = β −1 lnhe−βW i .


To interpret this result further it is useful to rewrite it further in terms of the cumulant expansion (see Chapter 1). This operation leads to the expression ∞ X 1 κn (−β)n−1 ∆F = n! n=1


where the first four cumulants are given by 2 κ1 = hW i , κ2 = h(W − hW i)2 i = σW


Nonequilibrium Statistical Mechanics


and 4 κ3 = h(W − hW i)3 i , κ4 = h(W − hW i)4 i − 3σW .


These expressions are instructive since we can infer from them different levels of approximation which are testable against experimental measurements or simulation results. To begin, we keep only the first term of the cumulant expansion. We then estimate the work done by the system as the free energy difference. This is correct only when the work is done reversibly, which means that there is no mechanism of energy dissipation in the system. The second level approximation amounts to consider 1 2 , ∆F = hW i − βσW 2


i.e., the fluctuations around the stationary state. Eq.(2.80) becomes exact in a regime near an equilibrium state, since then the work distribution is Gaussian, and all higher cumulants vanish identically. The result is also an example application of the fluctuation-dissipation theorem, since 1 2 βσ , 2 W relates the dissipated work to the Gaussian fluctuations. W diss =


If we want to take the full Jarzynski result serious, we have to estimate the free energy difference by the following expression " # N 1 X −βWi 1 e , (2.82) ∆F = − ln β N i=1 where N is the number of trajectories for which the work W has been determined. What trajectories are best to measure in an experiment or a simulation? This can be understood from reconsidering the Jarzynski-result in the form involving the dissipated work. Since we are near an equilibrium state, W is Gaussian-distributed, hence the dissipated work follows a Gaussian distribu2 2 tion with mean W diss = 12 βσW and variance σW . As W diss increases - one moves away from the near-equilibrium regime - the distribution will broaden. The expression eq.(2.82) obviously heavily weighs those trajectories whose dissipated work is negative. The probability of finding such a trajectory is given by Z 0 q 1 (2.83) P (Wdiss < 0) = dWdiss P (Wdiss ) = (1 − erf ( W diss /2)) 2 −∞


Computational Biology: A Statistical Mechanics Perspective

where erf is the error function.9 Since eq.(2.83) is a sharply decreasing function of its argument, it is established that the efficiency of sampling falls rapidly with the increase of dissipated work. Jarzynski’s equality has been tested in experiments on pulling RNA secondary structure (for details on RNA secondary structure, see Part II of the book). In such experiments - using a similar experimental setup we described in Chapter 1 in the context of the Worm-Like Chain model - a single RNA molecule with known secondary structure is fixed at its extremeties between two beads which can be moved reversibly, at different speeds (J. Liphardt et al., 2002). These experiments have indeed allowed to demonstrate that the estimator given by eq.(2.82) converges towards ∆F if sufficiently many trajectories N are taken into account for which the dissipated work is high. The quantity can also be validated in simulations (S. Park et al., 2003). Again, a sufficient sampling range is needed. More recently, a detailed study of RNA folding free energies has been performed by D. Collin et al., 2005, verified a generalization of the Jarzynski theorem, the Crooks fluctuation theorem, in regimes near and far from equilibrium. In the context of RNA folding, this theorem states that the probability the ratio of the probability distributions of the work for unfolding and folding under conditions of time-reversal symmetry fulfills Punf old (W ) = exp β(W − ∆F ) . (2.84) Pf old (W )


The Fokker-Planck and Langevin equations

In this section, we discuss two further classic analytical approaches commonly used to describe the stochastic dynamics of particle systems. The Fokker-Planck equation. In this subsection we will introduce a continuous approximation to the master equation. First, we recall the last general expression we had obtained for it, eq.(2.24) Z ∂t P (y, t) = dy 0 [W (y|y 0 )P (y 0 , t) − W (y 0 |y)P (y, t)] . (2.85) 9 The

error function is defined by 2 erf (x) ≡ √ π


Z 0


dte−t .

Nonequilibrium Statistical Mechanics


We now rewrite the transition probabilities W as W (y|y 0 ) = W (y 0 ; r) ,

r = y − y0


and obtain Z ∂t P (y, t) =

Z drW (y − r ; r)P (y − r, t) − P (y, t)

drW (y ; −r) .


This is just a rewrite, but now we want to play with the difference variable r = y − y 0 which will allow us to formulate a continuum approximation. In order to formulate this continuum version of the master equation we assume for the dependence of the transition probabilities on r that they fulfill the following W (y 0 ; r) ≈ 0 ,

|r| > δ

W (y 0 + ∆y; r) ≈ W (y 0 ; r) ,

(2.88) (2.89) (2.90)

|∆y| < δ

which means that W is a slowly varying function of y 0 , but sharply peaked in its dependence on r. If, additionally, P varies also slowly with y, we can expand in a Taylor series to obtain Z Z ∂t P (y, t) = drW (y; r)P (y; t) − dr r ∂y [W (y; r)P (y, t)] (2.91) 1 + 2


dr r2 ∂y2 [W (y; r)P (y, t)] − P (y, t)

Z drW (y; −r)

where the first and the last term on the rhs cancel each other out. The remaining integrals over r can be absorbed in the definition of the jump moments Z ∞ aα = drrα W (y; r) (2.92) −∞

and we finally obtain 1 ∂t P (y, t) = −∂y [a1 (y)P (y, t)] + ∂y2 [a2 (y)P (y, t)] . 2


This is the Fokker-Planck equation. It contains a drift term ∼ ∂y P and a diffusion term, ∼ ∂y2 P . Remember that we had introduced this terminology before in the context of the general formulation of the dissipation-fluctuation theorem. The Fokker-Planck equation makes this relationship again explicit: if we suppose a stationary distribution, ∂t P = 0, drift and diffusion term have


Computational Biology: A Statistical Mechanics Perspective

to balance each other. Task. Rederive the dissipation-fluctuation theorem of Section 2.3 for the Fokker-Planck equation. Example for the Fokker-Planck equation: Brownian motion. Consider a particle suspended in a liquid. If we trace its motion under the influence of random molecular collisions of the liquid, we find a continuous path, as shown in Figure 2.2. Measuring the distances the particle travels, and averaging over several realizations we obtain for the jump moments the expressions h∆xiX h(∆x)2 iX a1 ≡ = 0 , a2 ≡ = const. (2.94) ∆t ∆t Thus, the Brownian particle obeys a diffusion equation ∂t P (x, t) =

a2 2 ∂ P (x, t) 2 x


and we can identify a2 /2 ≡ D, i.e. the diffusion coefficient is given by D=

h(∆x)2 i . ∆t


We now want take a second, different look at the Brownian particle.10 Let us now consider the velocity instead of the position of a suspended particle as the dynamic variable. We first ignore the random collisions of the particle of the solution molecules and assume that the velocity of the particle relaxes according to v˙ = −γv ,


i.e., the particle velocity will go to zero for t → ∞: the differential equation is the same as that for the mean number of proteins in the protein degradation problem we discussed with help of the master equation. Hence, there is a now a drift term in the corresponding Fokker-Planck equation, and it is given by the jump moment 10 What we consider here is actually called the Rayleigh particle which is equivalent to the Brownian particle; the difference between the two is the fine graining of the time scale. In the discussion of the Rayleigh particle one assumes that tcoll  ∆t  trelax , where tcoll is a molecular collision time, and trelax the time scale on which the particle velocity relaxes.

Nonequilibrium Statistical Mechanics


FIGURE I.2.2: Particle trajectory in solution: a Brownian path.

a1 (v) =

h∆viV = −γv . ∆t


The second jump moment fulfills a2 (v) > 0 even for v = 0. The Fokker-Planck equation reads as ∂t P (v, t) = γ∂v [vP (v, t)] +

a2 2 ∂ P (v, t) . 2 v


The stationary distribution for this process is actually known from equilibrium statistical mechanics: it is the Maxwell-Boltzmann velocity distribution in which m is the particle mass and β = 1/kB T ,  P (v) =

βm 2π


  βmv 2 exp − . 2


With the help of this expression, we can identify the coefficient of the diffusion term as a2 /2 = γ/(mβ), and the Fokker-Planck equation is fully given by   1 ∂t P (v, t) = γ∂v vP (v, t) + ∂v P (v, t) . (2.101) βm


Computational Biology: A Statistical Mechanics Perspective

Exercise. Compute hv(t)i and hv 2 (t)i from eq.(2.101) for the given initial velocity v(0) = v0 . Exercise. Compute the dissipation-fluctuation relation for the Brownian motion from the Fokker-Planck equation for P (v, t). The Langevin equation. We now complete the discussion of the Brownian motion of the suspended particle in the velocity description by the classic approach originally suggested by P. Langevin. In order to do this we consider the velocity process under the inclusion of a noise source η(t) which we add to the rhs of equation (2.97), v˙ = −γv + η(t) .


The noise source η models the random action of the solution molecules on the suspended particle (we now put m = 1). The solution particles ‘kick’ the suspended particle and transfer part of their thermal energy. For the first two moments of the noise distribution we assume hη(t)i = 0 , hη(t)η(t0 )i = Γδ(t − t0 ) ,


where Γ is a constant. The motivating idea behind this specific choice is that the random collisions are instantaneous and uncorrelated. Exercise. How would the rhs of the second moment change if one were to introduce a small but finite collision time τc ? If we assume an initial velocity v(0) = v0 of the particle, as we did before, we can compute the velocity v(t) as v(t) = v0 e −γt + e−γt




dteγt η(t0 )



and, using the moments of η, we can obtain those of v as hvi = v0 e −γt hv 2 (t)i = v02 e −2γt +


Γ (1 − e−2γt ) . 2γ


We are now left to determine the coefficient Γ. For t → ∞, we know from eq.(2.100) - for m = 1 - that hv 2 (t → ∞)i = kB T =

Γ . 2γ


Nonequilibrium Statistical Mechanics


Again we have found an example of the dissipation-fluctuation theorem. A further, very prominent example is the Einstein relation D=

kB T , γ


see the Exercise on the Fokker-Planck equation in velocity space. Fokker-Planck vs. Langevin. The Fokker-Planck equation determines the full stochastic process of the Brownian particle; by contrast, the Langevin equation does, by construction, not go beyond second moments. Therefore, if we additionally assume that the noise η is Gaussian, all odd moments will vanish, and the even moments will factorize (Exercise!). This leads to what is commonly called a Gaussian white noise, i.e., a noise spectrum containing all frequencies, which usually serves to model a rapidly fluctuating force. This construction is all fine for the simple case of a Brownian particle, and lends itself to the many applications of the Langevin equation approach, in which a Gaussian white noise is added to the known deterministic equations of the system. The constant Γ is then usually adjusted such that the stationary solution correctly matches with the fluctuations around the stationary state. But it is worth keeping in mind that this approach is a strong simplification, as has been advocated by N. van Kampen. We will briefly go through the main points here; for a more detailed discussion the reader is referred to van Kampen’s book. Furthermore, we will return to this discussion at the end of Part III when we will discuss the role of fluctuations in biological systems. What has to be kept in mind if one wants to treat stochastic fluctuations with the Langevin approach? By this we refer to the procedure we used for the description of Brownian motion, namely to first write down a deterministic equation for the macroscopic (average) dynamics, and to then add on the fluctuations. i) Suppose your system is described by a deterministic equation of the type u˙ = A(u) = sin u to which we add the noise as defined above, ∂t u = sin u + η(t) .


If we average this equation, we find ∂t hui + hsin ui = 0


since, as before, we have hη(t)i = 0. This result, however, means that the


Computational Biology: A Statistical Mechanics Perspective

average does not obey the macroscopic equation, i.e., ∂t hui 6= sinhui, since 1 hsin ui = sinhui − h(u − hui)2 i coshui + ... , 2


which is an equation involving all higher moments. The message of this calculation obviously is: if we start from a deterministic equation for the average which is nonlinear, the simple addition of fluctuating source will in general be too naive. ii) For an arbitrary nonlinearity A(u) the Langevin equation eq.(2.109) is equivalent to the Fokker-Planck equation ∂t P (u, t) = −∂u (A(u)P (u, t)) +

Γ 2 ∂ P (u, t) . 2 u


If we allow equations of the type ∂t u = A(u) + C(u)η(t)


in which a u-dependent function multiplies the noise, we run into an interpretation problem. For each jump in the solution u of the equation due to the noise, the value of u and hence of C(u) is undetermined. We thus have to specify a rule how to interpret the product C(u)η(t), and this leads to a dependence of the resulting Fokker-Planck equation on that rule. Some possibilities are to take the value of C(u) before the jump, after the jump or the mean. The different options lead to different Fokker-Planck equations and hence to different results.11 The option to take the mean value is named after R. L. Stratonovich, the version to take the value before the jump is named ˆ. after K. Ito We illustrate this phenomenon for the example of protein degradation for which we had the master equation p˙n = γ(n + 1)pn+1 (t) − γnpn (t) . Within the Langevin approach we can assume √ n˙ = −γn + nη(t)



since we expect the fluctuations in the decay process to be proportional to the square root of the number of proteins. Thus, we have a case in which we 11 The

two stochastic equations differ by what is called a spurious drift term, and the equations can be transformed into each other by corresponding transformation rules; see the detailed discussion in van Kampen’s book.

Nonequilibrium Statistical Mechanics


need to determine the multiplication rule for the noise. In the Itˆ o-case, we take the value before the jump, hence n(t + ∆t) − n(t) = −γn(t)∆t +

Z n


dt0 η(t0 ) .



It can be proved that this choice leads to the Fokker-Planck equation ∂t P (n, t) = ∂n (γ +

Γ ∂n )[nP (n, t)] . 2

The corresponding line of reasoning in the Stratonovich case yields    Γ Γ P + ∂n [nP (n, t)] . ∂t P (n, t) = ∂n γn − 2 2



If we calculate the equation for the average hni ˙ from the Itˆo-version of the Fokker-Planck equation, we obtain the same result as eq.(2.28) which we obtained directly from the master equation (Exercise). The Stratonovich equation, by contrast, does not yield this result. The reason is that in the process of protein degradation, the probability to go from n to n − 1 proteins is indeed proportional to the number n of available proteins before the transition, in line with the Itˆ o assumption of the construction of the stochastic process in the first place. The ambiguity arising from the different possible choices is referred to as the Itˆ o-Stratonovich dilemma. A better understanding of what is an adequate procedure in formulating stochastic equations can be obtained from a distinction between intrinsic and extrinsic noise sources. Extrinsic noise is an ‘add-on’ to a fundamentally deterministic dynamics, as it is the case in many engineering applications. Here, one models, e.g., an electrical circuit on the level of its macroscopic constituents like resistors and capacitors and not on the level of electrons. In such a case, noise is never really white but has a finite correlation time. In this case, there is no ambiguity and the Stratonovich prescription applies. Intrinsic noise, by contrast, is due to the fact that the system itself is composed of discrete particles which interact stochastically: this noise can never be switched off, and hence A(u) is not determined by the evolution equations in a system isolated from noise. Hence, the Langevin approach does in general not work for internal noise. Instead, one has to go back to the master equation approach, from which in certain cases macroscopic equations can be derived.


Computational Biology: A Statistical Mechanics Perspective

This derivation can be done as follows in a systematic way. The starting point is the probability density P (y, t) for a stochastic process YX . We define Z y(t) ≡ hY it = dy y P (y, t) (2.119) where we assume that the density P is a sharply peaked function of its argument y; it is assumed to be initially of the form P (y, 0) = δ(y − y0 ). As t increases, the probability density P will evolve along y, and this determines the behaviour of the average y(t). One finds Z y(t) ˙ = dy y ∂t P (y, t) Z = dy dy 0 y [W (y|y 0 )P (y 0 , t) − W (y 0 |y)P (y|t)] (2.120) Z = dy dy 0 (y 0 − y)W (y 0 |y)P (y, t) . (2.121)

We now recall the jump moments (eq.(2.92)) Z aα (y) = dy 0 (y 0 − y)α W (y 0 |y)


for α = 0, 1, 2, .., to obtain d y(t) ˙ ≡ hY it = dt

Z dy a1 (y)P (y, t) = ha1 (Y )it


If we expand a1 (y) around the average hY it in a Taylor series 1 ha1 (Y )it = a1 (hY it ) + h(Y − hY it )2 it a001 (hY it ) + ... 2


one finds to a first approximation the equation y(t) ˙ = a1 (y(t)) = A(y)


which is, despite P fulfilling a linear equation, in general a nonlinear equation of y(t), even when the higher order terms are dropped. Task. Compute the corrections to the macroscopic equations. For this, consider d 2 hY (t)it (2.126) dt

Nonequilibrium Statistical Mechanics


and the variance σ 2 (t) = hY 2 (t)i − y 2 (t). For jump moments a1 and a2 which are linear in y show that the macroscopic equation can be written as 1 y(t) ˙ = a1 (y(t)) + σ 2 (t)a001 (y(t)) . 2


This concludes our discussion of the general methods of stochastic dynamics, i.e., master equations, Fokker-Planck and Langevin equations. We now move on to a class of systems to which these approaches will be applied later, to biochemical reaction systems, and give the basic elements of the theory of chemical kinetics. We discuss two approaches: first, the deterministic one, based on rate equations. Subsequently, we return to the stochastic aspect by a description of the Gillespie algorithm which allows to simulate the master equation for biochemical reactions. Chemical kinetics. In many applications of stochastic processes in biology, we have to deal with biochemical reactions - we will see such examples in Part III of the book. Consider a system of m chemical substances, and n chemical reactions. We represent the chemical reactions they undergo by n collision diagrams, given by νi1 A1 + ... + νim Am →ki µi1 A1 + ... + µim Am


for i = 1, .., n. Ai can represent either the number of a molecule of substance i, or its concentration (obtained by division through volume, ni = Ai /Ω). The ki are rate constants; νim , µim are stochiometric coefficients, which are integer numbers. The sums m X ri = νij (2.129) j=1

define the order of the transitions. A transition A1 + A2 →k A3


thus represents the binary collision of A1 and A2 , giving rise to A3 . Higher order collisions have a low probability to occur, and are generally neglected in chemical kinetics. In order to pass from a collision diagram to equations describing the time evolution, several assumptions have to be made. We begin with the following: • the system (typically, a solution) can be assumed spatially homogeneous;


Computational Biology: A Statistical Mechanics Perspective • the density of the molecules is low;12 • all reactions occur at a constant volume and temperature; • the collision frequency of the molecules depends on local concentration.

We can then write down the equation for the chemical kinetics of the molecules in the form n X dAj = (2.131) ki (µij − νij )Aν1i1 · · · Aνmim dt i=1 for j = 1, ..., m. This expression is the rate equation. The equations expressed in (2.131) are not always independent. This can be made more explicit by considering them in matrix form d A=M·K dt


where Aj is a vector with components Aj , M is a matrix with row vectors VT ≡ (µ1j − ν1j , ..., µnj − νnj ) for j = 1, ..., m, and KT ≡ (k1 Aν1i1 · · · Aνm1m , ..., kn Aνnn1 · · · Aνmnm ). If the matrix M has a rank r with r ≤ min{n, m} the equations are linearly dependent; there are thus m − r conservation laws which can be written in the form m X

αjk Aj (t) =


m X

αjk Aj (0)



for k = 1, ..., m − r. Task. Consider eq.(2.131) for the case of one collision diagram, with reaction and back-reaction rate constants k+ and k− . a) Write down the chemical master equation using the step operator E. b) For an ideal mixture of molecules, the grand canonical distribution is given by Y (Ωzj )nj P ({ni }) = exp(−Ωzj ) (2.134) nj ! j with nj = 0, 1, .... Here  zj =

12 But

2πm β

3/2 X

exp(−βν )



not too low; if the molecules appear in too few numbers, the continuum approach breaks down.

Nonequilibrium Statistical Mechanics


is the partition function of one molecule j in a unit volume. The energies ν contain all internal molecular degrees of freedom, be they rotational, vibrational or electronic. Check that P ({ni }) is a stationary solution of the master equation obtained under a), if Y µ −ν k+ = zj j j k− j


applies, i.e., the law of mass action is fulfilled. We will turn to the discussion of specific examples of kinetic equations in Part III of the book. The Gillespie algorithm. Since master equations are usually easy to write down but hard to solve explicitly, D. T. Gillespie, 1977, proposed a simple prescription how to determine the time evolution corresponding to the master equation from a stochastic algorithm. Starting point of the algorithm is the expression for the transition probability of the Chapman-Kolmogorov equation, eq.(2.21), in the case of small time interval τ . The algorithm can be formulated in terms of the probability P for a given reaction i = 1, ..., n to occur in an infinitesimal time-interval dτ , which can be written as P (τ, i) = P0 (τ ) · ai · dτ


where P0 (τ ) is the probability that no collision will have occurred in the interval (t, t + τ ), while ai ≡ hi ci is a stochastic reaction rate ci , multiplied by a combinatorial factor determined by the type of reaction that occurs, and by counting the number of distinct reaction partners that are available for a reaction. Thus, ai · dτ is the probability that a collision will occur in the interval (t + τ, t + τ + dτ ). Defining a0 ≡


ai ,



P0 (τ ) obeys the equation P0 (τ + dτ ) = P0 (τ ) · [1 − a0 · dτ ]


which describes the (obvious) fact that the probability that no reaction occurred in time τ +dτ equals the product of the probabilities that no transition


Computational Biology: A Statistical Mechanics Perspective

occured in the interval τ , and within dτ . The solution of this equation is given by P0 (τ ) = exp(−a0 τ ), such that P (τ ; i) = ai exp(−a0 τ ) , 0 ≤ τ < ∞ , i = 1, .., n,


or 0 otherwise. This result can be used to compute the time evolution of the system by the reconstruction of P (τ ; i) from a draw of two random numbers r1 and r2 from the unit interval uniform distibution according to the prescription   1 1 τ= ln , i = r2 (2.141) a0 r1 where i is selected such that the condition i−1 X

aj < r2 a0 ≤


i X




is fulfilled. These steps can finally be cast into the following algorithm: • define the stochastic rates ci for i = 1, ..., n, and the initial conditions on the N molecules; Pn • calculate a1 = h1 c1 , ..., an = hn cn , and a0 = j=1 aj ; • generate r1 , r2 and calculate τ and i; • increase t by τ , adjust the population levels of the selected reaction i; loop.


Sequence alignment: a nonequilibrium phase transition

Now we are fully done with the basic methods of stochastic processes. In the last section of Part I of this book we will now discuss an example which draws from all aspects that were discussed in the two Chapters of this Part: we will discuss a phase transition, but not an equilibrium one, we will find critical exponents and scaling, and we will both use stochastic equations in their discrete and continuous versions. The topic is sequence alignment, a

Nonequilibrium Statistical Mechanics


fundamental method in computational biology developed to find out the similarities between two sequences, be they made up DNA bases, or the amino acid sequences of proteins. In this section, we will see that the alignment of two sequences can be understood as a phase transition in a nonequilibrium system. This may, at first sight, seem astonishing: what does a pattern matching problem have to do at all with a phase transition? There are several aspects that have to be addressed in order to answer this question, and we will do so as we go along, and follow the exposition by R. Bundschuh, 2002. We begin with a technical definition of the notion of an alignment. Gapless alignment. The simplest procedure of sequence alignment is called gapless alignment. It looks for similarities between two sequences a = (a1 , a2 , ..., aM ) and b = (b1 , b2 , ...., bN ) where M ∼ N . The letters ai , bi are taken from an alphabet of size c; for our purposes here we take c = 4, i.e., the four-letter alphabet of the DNA bases. A local gapless alignment A consists of two substrings of equal length ` of the sequences a, b. To each such alignment can be assigned a score S[A] = S(i, j, `) =

`−1 X

sai−k ,bj−k



where the scoring matrix is given by in the simplest case  1 a=b sa,b = . −µ a 6= b


What is to be computed are the values of i, j and ` which lead to the highest total score Σ ≡ max S[A] (2.145) A

for the given scoring matrix sa,b . This optimization problem can be reformulated by introducing the auxiliary quantity Si,j which is the optimal score of the subsequences ending at (i, j) optimized over `. This quantity can be computed with O(N 2 ) steps instead of O(N 3 ) with the prescription Si,j = max{Si−1,j−1 + sai ,bj , 0}


for the initial condition S0,k = Sk,0 = 0. This recursion expresses the fact that for a given pair (i, j) the optimal ` = 0 or ` > 0. If ` = 0, the resulting score is zero either; if ` = 1 at least, the corresponding pair (ai , bj ) will belong


Computational Biology: A Statistical Mechanics Perspective

to the optimal alignment (which may be longer), whatever had been chosen as optimal up to (i − 1, j − 1). Eq.(2.146) describes a random walk with increment sa,b . It is cut off when it falls below zero. The global score Σ is then given by Σ= max Si,j . (2.147) 1≤i≤M,1≤j≤N

Significance of the alignment. Suppose we have found an alignment by performing this computation. How significant is it? In order to answer this question, we have to discuss the alignment of purely random sequences, and to determine the distribution of scores in this case. This can be done rigorously, leading to the result P [Σ < S] = exp(−κe −λS ) ,


which is a Gumbel or extreme value distribution. It is characterized by two parameters λ and κ where λ characterizes the tail of the distribution. For the case of gapless alignment we discuss here, both parameters can be calculated from the knowledge of the scoring matrix sa,b . We want to illustrate how the result (2.148) comes about in a heuristic fashion. For this we go back to equation (2.146) and set i = j, which is permissible for random sequences. Thus we have Si,i ≡ S(i) = max{S(i − 1) + s(i), 0} .


In this equation, s(i) = sa,b plays the role of an uncorrelated noise given by the distribution X P [s(i) > s] = pa pb . (2.150) (a,b|sa,b >s)

Eq.(2.149) is essentially a discrete version of a Langevin equation. The dynamics it generates can be found to be in two distinct phases, called the linear and the logarithmic phase. The quantity which distinguishes the two is the local similarity score X hsi = pa pb . (2.151) (a,b)

In the linear phase, the dynamics P is a random walk with an average upward drift hsi, and the maximal score is ≈ N hsi for a sequence of length N . This phase constitutes a phase of global alignment, and hence does not permit to identify similarities in subsequences. The distribution of the Σ in this phase is Gaussian, not Gumbel. If, however, the average drift hsi is negative, the ensuing dynamics will lead to a score of the form shown in Figure 2.3. The resulting score landscape can be considered as consisting of ‘islands’ in an

Nonequilibrium Statistical Mechanics


FIGURE I.2.3: Total score of an alignment as a function of sequence position (after R. Bundschuh, 2002) ‘ocean’. For the case of random sequences, the islands are statistically independent. If σk is the maximal score of island k, the σk are thus independent random variables. Island distribution. In order to calculate the island distribution explicitly, we have to look at (R. Bundschuh, 2002) p(σ) = hδ(σ −


s(i))i .



Here L is the length of a ‘large’ island, measured from its beginning to the peak at height σ. Using the Fourier representation of the δ-function and the statistical independence of the s(i) we obtain Z 1 dke −ikσ he iks iL (2.153) p(σ) = 2π Assuming that the peak score of each island is proportional to its length islands are thus on average little triangles with rising slope α - we approximate Z 1 dke −ikαL he iks iL (2.154) p(σ) ≈ 2π and then evaluate the integral in a saddle-point approximation. This leads to (Task!) p(σ) ∼ exp(−λσ) (2.155) with λ = iks − ln[he iks s i]/α ,


where the saddle-point value of k, which we call ks , is determined by hse iks s i = α. he iks s i



Computational Biology: A Statistical Mechanics Perspective

Note that ks still depends on α. The correct value α is found by minimizing eq.(2.157) with respect to α and using eq.(2.157), which yields heiks s i = 1




i.e., explicitly he λs i =

pa pb e λsa,b = 1 .


The typical slope of an island is given by α = hse λs i .


Thus, we conclude that the islands follow the exponential distribution P [σk > σ] ≈ Ce −λσ .


Since the global optimal score Σ is given in terms of the islands σk by Σ = max{σk } k


the distribution of the Σ can be computed from the distribution of the σk . For a large number K ∼ N of island peaks one finds P [Σ < S] = P [max{σ1 , .., σK } < S] = [1 − Ce −λS ]K ≈ exp(−κe −λS ) , (2.163) where κ = CK. Gapped alignment. We now turn to the case of alignment with gaps. This method is used to detect weak sequence similarities between evolutionary distant sequences, in which deletions and insertions have occured over time. The classic example is Smith-Waterman local alignment, in which the two subsequences a, b, e.g. GATGC and GCTC, can be aligned as GATGC and GCT-C, i.e., with one gap (see Figure 2.4). The score function for alignment with gaps is given by X S[A] = sa,b − δNg , (2.164) a,b

where Ng is the number of gaps with cost δ. The gapped alignment of two sequences can be represented as a directed path on a two-dimensional lattice, see Figure 2.4. The alignment score is the sum over local scores of the traversed bonds, whereby diagonal bonds are gaps with penalty δ, while horizontal bonds are given the similarity scores s(r, t) ≡ sai ,bj . What is sought for is the best scoring path connecting the lattice origin (0, 0) to its end, (0, 2N ).

Nonequilibrium Statistical Mechanics


FIGURE I.2.4: The local alignment of the two sequences CGATGCT and TGCTCGA is represented as a directed path on an alignment lattice: the diagonal bonds correspond to gaps, while horizontal bonds are aligned pairs (Reprinted with permission from R. Bundschuh, 2002. Copyright: American Physical Society.)

If we denote by h(r, t) the score of the best path ending in a lattice point h(r, t), the highest scoring global alignment can be computed with the Needleman-Wunsch algorithm h(r, t + 1) = max{h(r, t − 1) + s(r, t), h(r + 1, t) − δ, h(r − 1, t) − δ} . (2.165) This expression can be interpreted to describe the configuration of a directed polymer in a random potential given by the local scores s(r, t). Another interpretation, which is somewhat easier to visualize, is to understand the configuration as the height profile h(r, t) of a growing interface between a solid and, say, a liquid phase. These systems - directed polymers in a random potential or a growing interface - are well-known in the physics of nonequilibrium systems. Due to the mapping of the sequence alignment problem onto the interface growth problem one can immediately show, from the knowledge of the models in nonequilibrium physics, that it belongs to the so-called KPZ universality class. A universality class comprises all model systems that can be shown to be mathematically equivalent and will have the same phase transition - i.e., they can


Computational Biology: A Statistical Mechanics Perspective

be characterized by a common set of critical exponents. Within a continuum description, one can show that the evolution of the heights is governed by the Kardar-Parisi-Zhang equation (often short: KPZ equation) (M. Kardar, G. Parisi and Y.-C. Zhang, 1986; T. Hwa and ¨ssig, 1998) M. La ∂t h = ν0 ∂r2 h + λ0 (∂r h)2 + η(r, t)


where η(r, t) = 21 s(r, t)−δ is an uncorrelated Gaussian white noise. For t → ∞ the heights assume a stationary equal-time distribution ! 1 X 2 P [h(r, t → ∞)] ∝ exp − [h(r + 1, t) − h(r, t)] . (2.167) 2D r The parameter D in this expression is a function of the scoring parameters µ and δ. Typical score profiles h(r, t) are illustrated in Figure 2.5 (top). In order to extract the characteristics of these profiles, it is instructive to consider the width of the profile, w(t), defined by w2 (t) =

1 X



[h(x, t) − h(t)]2



with the interval size X(≈ N ), and where the spatial average of the height, h(t) is defined by x=X/2 1 X h(t) = h(x, t) . (2.169) X x=−X/2

The behaviour of h and w is shown in Figure 2.5 (bottom). As illustrated in the Figure, the asymptotic behaviour of the width obeys a scaling law w(t) = B(µ, δ)tω


where the value of ω is known from the universal scaling behaviour of the KPZ equation. Note that the scaling behaviour is obtained from an ensemble average over different realizations of random sequences. The quantity ∆h(t) = hmax (t) − hmin (t) displays the same scaling behaviour as w.


Nonequilibrium Statistical Mechanics


FIGURE I.2.5: Score profiles h(x, t) (top) and scaling result (bottom) for ¨ssig, 1998). ∆h(t) and w(t) in a log-log plot (after T. Hwa and M. La


Computational Biology: A Statistical Mechanics Perspective

FIGURE I.2.6: Score profiles h(x, t) (top) for weakly correlated sequences, ¨ssig, and scaling result for ∆h(t) in a log-log plot (after T. Hwa and M. La 1998).

Nonequilibrium Statistical Mechanics


FIGURE I.2.7: The phase diagram of the log-linear phase transition in the parameters δ vs. µ. The variable c the number of letters in the alphabet; a value of 4 hence refers to the alphabet of DNA. (Reprinted with permission from R. Bundschuh, 2002. Copyright: American Physical Society.) If the sequences exhibit correlations, the behaviour changes. This is illustrated in Figure 2.6, which displays the growth of a peak in the score profiles. In terms of the height difference ∆h(t) one sees a deviation from the scaling behaviour of the random case, from a power law of the form t 1/3 to a linear law. This change in exponent illustrates the onset of global alignment, and we are thus back to the linear phase we discussed before for the case of gapless alignment. As a final step we want to locate the logarithmic phase in parameter space. For the gapped case, a negative drift hsi is not sufficient, since the average score now has to grow by a gap-dependent amount u({sa,b }, δ) on top of the expectation value hsi. Consequently, the log-linear transition occurs at u({sa,b }, δ) + hsi = 0 .


In a (δ, µ) diagram this condition corresponds to a line δc (µ), which can be calculated approximately or numerically. The result is shown in Figure 2.7.


Computational Biology: A Statistical Mechanics Perspective Additional Notes

A basic reference for the description of nonequilibrium processes in physics and chemistry is the book by N. van Kampen, 1992, which served as a basis for the discussion presented here. Readers willing to learn more details are thus recommended to continue further studies there. We have treated chemical kinetics in a very brief manner; for a much more detailed exposition, in particular on the kinetics of enzymatic reactions, readers are asked to consult the book by A. Cornish-Bowden, M. Jamin and V. Saks, 2005. The relation between equilibrium and nonequilibrium physics as expressed by the Jarzynski theorem is an active field of research. In particular the recent papers by C. Jarzynski, G. Crooks, C. Maes and U. Seifert are recommended reading for those who want to know more on this topic - consultation of the current literature is, however, required. On the experimental side, there are continuous efforts to validate the fluctuation theorems. Three recent contributions are by D. Keller et al., 2003, E. H. Trepagnier et al., 2004, and D. Collin et al., 2005. The literature on sequence alignment is (fairly obviously) extensive. A point of reference is the book by M. S. Waterman, 1995. The classic paper on alignment is by S. B. Needleman and C. D. Wunsch, 1970. For a detailed description of the development until 1995 the reader is asked to consult the Sources and Perspectives section in Waterman’s book. The relationship between the alignment problem and models from nonequi¨ssig, 1996 librium statistical mechanics was observed by T. Hwa and M. La and 1998. More recently, a mapping to the asymmetric exclusion process has been proposed by R. Bundschuh, 2002, which is the same universality class as the KPZ equation discussed in the text. The ASEP is a favorite model for studies of nonequilibrium statistical mechanics since it is amenable to rigorous approaches.

Nonequilibrium Statistical Mechanics


References R. Bundschuh, Asymmetric exclusion process and extremal statistics of random sequences, Phys. Rev. E 65, 031911 (2002) D. Collin, F. Ritort, C. Jarzynski, S. B. Smith, I. Tinoco Jr. and C. Bustamante, Verification of the Crooks fluctuation theorem and recovery of RNA folding free energies, Nature 437, 231-234 (2005) A. Cornish-Bowden, M. Jamin and V. Saks, Cin´etique Enzymatique, EDP Sciences (2005) G. E. Crooks, Path-ensemble averages in systems driven far from equilibrium, Phys. Rev. E 61, 2361-2366 (2000) D. T. Gillespie, Exact Stochastic Simulation of Coupled Chemical Reactions, J. Phys. Chem. 81, 2340-2361 (1977) J. Gore, F. Ritort and C. Bustamante, Bias and error in estimates of equilibrium free-energy differences from nonequilibrium measurements, Proc. Natl. Acad. Sci. USA 100, 12564-12569 (2003) T. Hwa and M. L¨ assig, Similarity Detection and Localization, Phys. Rev. Lett. 76, 2591-2594 (1996) T. Hwa and M. L¨ assig, Optimal Detection of Sequence Similarity by Local Alignment, in Proceedings of the Second Annual Conference on Computational Molecular Biology, S. Istrail (ed.), ACM Press (1998) C. Jarzynski, Nonequilibrium Equality for Free Energy Differences, Phys. Rev. Lett. 78, 2690-2693 (1997) C. Jarzynski, Equilibrium free-energy differences from nonequilibrium measurements: A master-equation approach, Phys. Rev. E 56, 5018-5035 (1997) M. Kardar, G. Parisi and Y.-C. Zhang, Dynamic Scaling of Growing Interfaces, Phys. Rev. Lett. 56, 889-892 (1986) D. Keller, D. Swigon and C. Bustamante, Relating Single-Molecule Measurements to Thermodynamics, Biophys. J. 84, 733-738 (2003) P. Langevin, Sur la th´eorie du mouvement brownien, C. R. Acad. Sci. (Paris) 146, 530-533 (1908)


Computational Biology: A Statistical Mechanics Perspective

M. Lax, Fluctuations from the Nonequilibrium Steady State, Rev. Mod. Phys. 32, 25-47 (1960) J. Liphardt, S. Dumont, S. B. Smith, I. Tinoco Jr. and C. Bustamante, Equilibrium Information from Nonequilibrium Measurements in an Experimental Test of Jarzynski’s Equality, Science 296, 1832-1835 (2002) C. Maes, On the origin and use of fluctuation relations for the entropy, S´em. Poincar´e 2, 29-62 (2003) S. B. Needleman and C. D. Wunsch, A general method applicable to the search for similarities in the amino acid sequence of two proteins, J. Mol. Biol. 48, 443-453 (1970) S. Park, F. Khalili-Araghi, E. Tajkhorshid and K. Schulten, Free energy calculation from steered molecular dynamics simulations using Jarzynski’s equality, J. Chem. Phys. 119, 3559-3566 (2003) U. Seifert, Fluctuation theorem for birth-death or chemical master equations with time-dependent rates, J. Phys. A: Math. Gen. 37, L517-L521 (2004) U. Seifert, Fluctuation theorem for a single enzym or molecular motor, Europhys. Lett. 70, 36-41 (2005) E. H. Trepagnier, C. Jarzynski, F. Ritort, G. E. Crooks, C. J. Bustamante and J. Liphardt, Experimental test of Hatano and Sasa’s nonequilibrium steady-state equality, Proc. Natl. Acd. Sci. USA 101, 15038-15041 (2004) N. G. van Kampen, Stochastic Processes in Physics and Chemistry, Elsevier (1992) M. Waterman, Introduction to Computational Biology, Chapman & Hall/ CRC (1995)

Part II


Chapter 1 Molecules, Code and Representation

The proper term for such a translation rule is, strictly speaking, not a code but a cipher. [...] I did not know this at the time, which was fortunate because “genetic code” sounds a lot more intriguing than “genetic cipher”.

Francis Crick, What Mad Pursuit (1988) In this Chapter, we give a brief list of the basic chemical properties of the biomolecules whose physical properties will occupy us in the following chapters. The details given here are minimal; the indispensable reference to check is the book by B. Alberts et al. (2002).


DNA and RNA: the building blocks

DNA and RNA are charged polymers composed of three structural elements, a sugar, a phosphate group and the bases. The sugar gives the molecules their names: deoxyribose or ribose. Both are pentose rings with five carbon atoms; the arrangement of the carbons and their numbering is indicated in Figure 1.1. The chemical formula of deoxyribose is C5 H10 O4 , of ribose C5 H10 O5 . The sugar builds, together with the phosphates, the backbone of DNA and

FIGURE II.1.1: The sugar ring of the backbone of both DNA and RNA. It contains an oxygen atom; the numbering of the carbon atoms is indicated.



Computational Biology: A Statistical Mechanics Perspective

FIGURE II.1.2: A DNA single strand with a thymine base offering two hydrogen bonds to bind to a complementary base, adenine.

RNA. The phosphates are attached to one sugar at the 5’-tail and to the next at the 3’-group, see Figure 1.2. Towards the opposite side, nucleotide bases are attached. Figure 1.3 displays the chemical structure of the four possible bases in DNA, two pyrimidines, and two purines, and the base uracil which replaces thymine in RNA. Neighbouring bases along a DNA strand experience stacking interactions when they are in registry, see Figure 1.4. Consequently, a single-stranded DNA or RNA molecule is characterised by its base sequence and the orientation of the strand, e.g. 50 − ACT GT T T T ACCCG − 30 . The strand orientation is from the 5’ to the 3’ prime end; this is called the sense strand, while the direction 3’ to 5’ is the anti-sense strand. The bases can provide hydrogen bonds to a complementary base with which it can hybridize to form a double strand. This base pairing mechanism gives rise to the doublehelical structure of the DNA molecule, see Figure 1.5. In RNA, the base thymine (T) is replaced by uracil (U), and the sugar is ribose instead of 2-deoxyribose. In an organism RNA typically comes as a single strand, since it is the product of the transcription (i.e., the reading process) of a gene by the readout-molecule RNA polymerase. (We will learn more about those processes in Part III of the book.) The basic variants of RNA are listed in Table 1.1 While DNA usually forms a double-strand. between two complementary

Molecules, Code and Representation


FIGURE II.1.3: DNA and RNA bases. Uracil replaces thymine in RNA.

FIGURE II.1.4: DNA and RNA stacking. Left: stacked bases, right: an unstacked base.


Computational Biology: A Statistical Mechanics Perspective

FIGURE II.1.5: The DNA double helix. (From the Talking Glossary of Genetics, US National Human Genome Research Institute.)

sequences, RNA frequently hybridizes with itself.1 As a consequence of selfhybridization, an RNA base sequence can give rise to a rich secondary structure, composed from RNA single strands interspersed with helical, i.e., bound, regions. These different structures correspond to the different functional roles the specific RNA molecule can play within its biological context. The most variable type is the messenger RNA (mRNA) which is synthesized as a transcript from the coding regions of DNA; it is the read-out of the transcription process by RNA polymerase. The fold of a transfer RNA is illustrated in Figure 1 DNA

can also self-hybridize, but this is the less frequent situation.

Molecules, Code and Representation


1.6. As its name says, transfer RNA has a transport function; it brings an mRNA transcript to the cellular bodies where protein synthesis occurs, the ribosomes. It has a characteristic cloverleaf structure, in which each of the leaves takes up a particular function.

FIGURE II.1.6: Transfer RNA, in a schematic representation.

The original understanding of RNA as being a purely passive intermediate on the way from the genetic code of DNA to the functional properties of proteins has obviously evolved in the past years, in particular through the discovery of the mechanism of RNA interference, in which small non-coding double-stranded RNA molecules can regulate gene expression (A. Fire et al., 1998). Apart from the classic variants, meanwhile a zoo of different RNA types has been found, some of which are indeed also double-stranded. RNA molecules are nowadays understood as being mostly non-coding, i.e., they do not help translate a DNA sequence into a protein. A list of the novel examples is provided in Table 1.2. One could be tempted to say that DNA is all information; by contrast, RNA is information and structure. DNA structure can vary between different


Computational Biology: A Statistical Mechanics Perspective

TABLE 1.1:

RNA variants

RNA type

size (nt)



variable 75-95 ≈ 103

transcript: template for protein adapter molecule part of ribosome; protein synthesis

TABLE 1.2:

Recent RNA variants.

sRNA small (non-coding) RNA, common name in bacteria, 75 − 400 bp ncRNA non-coding RNA, common name in eukaryotes miRNA microRNA, form a putative translational regulatory gene family stRNA small temporal RNA, often also called miRNA snRNA small nuclear RNA, includes eukaryote spliceosomal RNA snmRNA small non-mRNA, ncRNA snoRNA small nucleolar RNA, involved in rRNA modification tmRNA bacterial ncRNA, have both mRNA and tRNA function siRNA short (21-25 nt) interfering RNA, double-stranded, involved in gene silencing

conformational and topological states, but these are usually based on a doublestranded molecule. By contrast, the single-stranded RNA molecule can build a sequence-dependent three-dimensional structure, a fold, and in this way it acquires similar degrees of freedom as a protein. It is thus not sufficient to know the base sequence in order to characterize RNA: one also needs to know the fold in order to understand its function.


Representing RNA structure

The basic structural elements that occur in RNA secondary structure are summarized in Figure 1.7. This collection already indicates the complexity one encounters when one wants to classify and predict these structures in the context of a complete molecule. RNA structure is complex, but the number of the basic variable elements is only four, hence small when compared to the twenty amino acids of the proteins which we will encounter in next section. It is useful to briefly list the ideas used to represent the self-hybridized configurations of an RNA molecule. Two kinds of representations are shown in Figure 1.8.

Molecules, Code and Representation


The top graph of Figure 1.8 shows an RNA molecule represented explicitly by vertices (the nucleotides) and full-line edges. The edges are the connections between the vertices and represent the hydrogen bonds, shown as dotted lines; in the following, we drop this distinction. But there is also, different from Figure 1.7, a second class of edges, drawn in broken lines. Consider the difference between the two configurations, once without and once with the broken edges. Without the broken edges, the RNA could be cut in two without affecting the hybridized chains. This is not the case anymore when the broken edges are present: the RNA structure now has a pseudoknot, intertwining separate paired regions along the chain. This property becomes clear in the bottom graph. It can be thought to be obtained from the top graph by ‘pulling at the ends’ of the RNA chain. The arc or rainbow diagram clearly shows the ‘overlapping’ bonds for the pseudoknot. There are formal ways to represent these RNA configurations which can be used in computer implementations (see, e.g., G. Vernizzi et al., 2004). The first is a bracketing rule: an unpaired base is given by a dot, while a paired base is described by a bracket, indicating opening or closing of a bond. The top graph in Figure 1.7 would thus be represented as .(((.....)))....(((.....))).


for the case without pseudoknot, and as .(((...[[)))....(((]]...))).


when the pseudoknot-bonds are present. Each additional pseudoknot requires a new bracket-type to be introduced. Alternatively, one can represent the bonds along the chain also in an array or matrix form. For this one writes the sequence of L nucleotides as an L × L contact matrix C with elements Cij = 1, if i is paired to j, or zero otherwise. One can also interpret the pairing between any two bases i and j as a transposition of the elements {i, j} and associate a permutation structure to it via σ(i) = j if i, j are paired, and σ(i) = i if not. For the example sequence {50 −CU U CAU CAGGAAAU GAC −30 } we give the pseudoknotted structure in dot-bracket notation in the first row and the permutation structure in the second and third:   . ( ( ( . [ [ [ ) ) ) . . ] ] ] . σ =  1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17  (1.3) 1 11 10 9 5 16 15 14 4 3 2 12 13 8 7 6 17 This is an involution since σ 2 is the identity permutation.


Computational Biology: A Statistical Mechanics Perspective

FIGURE II.1.7: Schematic representation of the elements of RNA secondary structure; backbone and base pairs are both indicated as black bars, so no difference is made between covalent and hydrogen bonds. From top left to bottom right: single strand; duplex; duplex with dangling end; single nucleotide bulge; three nucleotide bulge; hairpin: stem (duplex region) and loop; mismatch pair or symmetrical loop; asymmetric internal loop; symmetric internal loop; two-stem junction; three-stem junction; four-stem junction.

Molecules, Code and Representation


FIGURE II.1.8: Graphical representation of RNA with pseudoknots (see text).

FIGURE II.1.9: Drawing RNA on a sphere and on a torus, after G. Vernizzi et al., 2004.


Computational Biology: A Statistical Mechanics Perspective

Finally, there is an even more fundamental way to look at RNA pseudoknots. If one closes the endpoints of a rainbow diagram to let it form a closed loop, one obtains a circle diagram. In this diagram, RNA structures with and without pseudoknots are distinguished by the presence or absence of edge crossings, when all of them are drawn either inside or outside the circle diagram. Exercise. Draw the circle diagram for the kissing hairpin, with and without the pseudoknot bonds, and with bonds lying either all inside or outside the circle. It is instructive to draw the circle diagram on the surface of a sphere, with the edges outside of the circle. Figure 1.9 (top) shows this for a nonpseudoknotted case. By contrast, a pseudoknotted configuration can be drawn without edge crossings on the surface of a torus (Figure 1.9, bottom). The relationship between RNA structures and topology can be made quantitative through the Euler characteristic of the surface on which they can be drawn without crossing the bonds. The Euler characteristic is defined as χ≡V −E+F


where V are the vertices (nucleotides), E the edges (bonds), and F the faces of closed loops. RNA chains without pseudoknots have χ = 1. For the kissing hairpin, one has χ = −1. This result can also be expressed in terms of the genus g of the surface, which is χ = 1 − 2g for the case at hand. It is then clear that the kissing hairpin can be represented by a torus, for which g = 1. We will come back to the topological classification of RNA pseudoknots in Chapter 2 of this Part.



Proteins are built-up from twenty amino acids; they are listed in Table 1.3 (following A. von Haeseler and D. Liebers, 2003). A triplet of DNA bases codes for an amino acid. Since the number of existing amino acids is lower than the combinatorial possibility based on the triplet rule, there is thus redundancy in the code. It appears typically, but not exclusively, in the last codon. Note that particular combinations of the bases also code for the start and stop of coding sequences. Predominantly hydrophobic amino acids are marked with an asterisk; no distinction as to their degree of hydrophobicity is made.

Molecules, Code and Representation

TABLE 1.3:

Codons and amino acids; s.c.: starter codon.








arginine serine lysine asparagine threonine methionine∗ /s.c. isoleucine∗ arginine glutamine histine proline∗ leucine tryptophane∗ stop codon cysteine∗ stop codon stop codon tyrosine∗ serine leucine∗ phenylalanine∗ glycine∗ glutamic acid aspartic acid alanine∗ starter codon (s.c.) valine∗

Arg Ser Lys Asn Thr Met Ile Arg Gln His Pro Leu Trp




Tyr Ser Leu Phe Gly Glu Asp Ala





Computational Biology: A Statistical Mechanics Perspective

FIGURE II.1.10: Protein structure. (From the Talking Glossary of Genetics, US National Human Genome Research Institute.)

Molecules, Code and Representation


Amino acids can further be distinguished by their hydrophobicity: hydrophobic amino acids will prefer to bury themselves inside of a protein fold in order to avoid water contact. The sequence of amino acids is called the protein primary structure. Proteins also form secondary structures which are recurrent structural elements. These are the α-helix, and the β-sheet; as the names indicate they correspond to a helical and a folded, but fairly planar structure. Like DNA, the molecular building elements form hydrogens bonds with each other to stabilize these structures. On the next level of complexity of proteins arises the full fold, the tertiary structure. Finally, fully folded proteins can form protein complexes or quaternary structures. These elements are all summarized in Figure 1.10.


Computational Biology: A Statistical Mechanics Perspective References

B. Alberts, A. Johnson, J. Lewis, M. Raff, K. Roberts and P. Walter, The Molecular Biology of the Cell, 4th ed., Garland Science (2002) A. Fire, D. Albertson, S. W. Harrison and D. G. Moerman, Production of antisense RNA leads to effective and specific inhibition of gene expression in C. elegans muscle, Development 113, 503-514 (1991) A. von Haeseler, D. Liebers, Molekulare Evolution, Fischer (2003) G. Vernizzi, H. Orland and A. Zee, Prediction of RNA pseudoknots by Monte Carlo simulations, preprint q-bio/045014 (2004)

Chapter 2 DNA

We have stated in the first Chapter that the most fundamental feature of DNA is its ability to hybridize two complementary strands. This is the prime example of the physico-chemical recognition processes which are at the heart of essentially all biological mechanisms on a molecular level. In some sense, the base-pairing mechanism of DNA is also the simplest of these recognition processes. It can be easily studied experimentally in vitro: one only needs to heat a solution containing double-stranded DNA molecules. The thermal denaturation or melting of DNA is the dissociation of the doublestranded DNA into its two separate single stands - hence just the inverse of the recognition mechanism. This process is reversible: by cooling the sample, the single-strands rehybridize to the double-strand. As first noticed by R. Thomas in the early 1950’s, this denaturation/hybridization process is indeed a thermodynamic phase transition in the limit of long chains. Many features of this phase transition remain correct in a more direct biological context, namely when the double strand of DNA is opened by the application of a direct force, e.g., by an enzyme. We will now start to describe the thermal denaturation process, and we begin with short chains. This case is simpler than that of longer chains and also has important applications in biotechnology, e.g., for DNA microarrays to which we will turn later on as well.


Thermal stability of DNA: the melting transition

Denaturing short chains. The opening of double-stranded DNA molecules has two aspects. Ignoring first any sequence-dependent effects, it can be understood as a simple dissociation process: as one double strand opens, it gives rise to two single strands. One can easily imagine that when the chains become longer and longer, the sequence-dependence will become more and more important, and dissociation will not simply be one double strand vs. two single strands anymore; many intermediate states will become possible in



Computational Biology: A Statistical Mechanics Perspective

which the double-stranded molecule is only partially dissociated. As we have said, in the case of short chains (typically up to, say, 50 bp) we want to model the denaturation process as a dissociation equilibrium C2 ↔ 2C1


where C1 is the concentration of single strands, and C2 the concentration of double-strands. This reaction is governed by the equilibrium constant KD = C2 /C12 . The total concentration of DNA is given by CT = C1 + 2C2 , since each double-strand contains two single strands. The quantities of interest are the fraction of double strands, θD ≡

2C2 . CT


or, likewise, the fraction of single strands, θ = 1 − θD . Using the definition of KD , a quadratic equation for C1 can be obtained which allows to express θD or θ in terms of KD and CT . The fraction of single strands is then found to fulfill the equation √ −1 + 1 + 2γ θ= , (2.3) γ where γ ≡ 4CT /KD . This calculation assumes differing strands; if the two strands are self-complementary, 4CT has to be replaced by CT . Why? Exercise! What remains to be calculated is the equilibrium constant KD ; in particular, its temperature dependence will turn out to be important. One has KD = exp[−(∆H − T ∆S)/RT ]


where the transition enthalpy is given by ∆H = ∆HH−bond + ∆Hnn


contains both contributions from the hydrogen bonds between the paired bases on the two strands, and from the stacking of the pairs on top of each other, see the previous section. Typically, for the stacking contribution it is assumed that it can be considered as a purely nearest-neighbor effect, but there are situations (e.g., in the presence of mismatches) when this is not sufficient.1 The transition entropy contribution in eq.(2.4) is typically approximated by ∆S = ∆Sbp [NAT + NCG ] where ∆Sbp is a temperature-independent value, 1 The

same philosophy applies to RNA-RNA and RNA-DNA-duplexes, but: the values of the stacking interactions differ in all these cases.



FIGURE II.2.1: Melting curve of a 16-bp oligomer, theory (line) and experiment. Insert: UV-absorption data. Reprinted with permission from J. Bayer et al.; Copyright (2005) American Chemical Society.

and NAT and NGC are the number of AT- and CG-base pairs, respectively. The precise values of the melting parameters have been determined experimentally, see, e.g., R. Owczarzy et al., 1998, and J. Santa-Lucia, 1998. Figure 2.1 illustrates the comparison between theory and experiment for the sequence 5’-TAG TTG TGA TGT ACA T-3’. (J. Bayer et al., 2005). The graph illustrates generic features observed in oligomer melting curves. The DNA double-strand undergoes a melting transition at a temperature TM which is at θ = 0.5; here, the slope of the curve is maximal (the derivative dθ/dT has a maximum). The melting temperature TM depends on base composition and strand length. Further, the shape of the melting curve turns out to be universal, it generically has a sigmoidal form. The effect of a change in base sequence thus leads to a shift in melting temperature while the overall curve is unchanged. If the base sequence is extended by adding more bases, the melting temperature will shift to higher temperatures - simply because more bases have to be broken. At the same time, however, the sigmoidal curve will steepen up. This is a consequence of a cooperative effect: as long as the denaturation process occurs in a two-state fashion, essentially all bases have to open up in concert.


Computational Biology: A Statistical Mechanics Perspective

Melting long chains: the Poland-Scheraga model. We now turn to the discussion of melting of long DNA chains which do not open in the simple two-state fashion. We can easily imagine that if we make the chains longer and longer, from 50 bp to, say, 500 bp, the cooperative effect of all bases opening in concert will not be operative any more all along the chain. We may then expect that a DNA molecule undergoing the denaturation process may look like the schematic illustration in Figure 2.2. The molecule consists of a sequence of helices and open segments, so-called denaturation loops or bubbles. The configuration shown in addition has open ends, but closed ends are of course possible, too.

FIGURE II.2.2: A DNA chain configuration near the melting transition: it consists of an open end of length `0 , a bound segment of length `1 , a loop of length `2 , a bound segment of length `3 , and an open end of length `4 .

A model to describe the statistics of these configurations was first proposed by Poland and Scheraga, 1970; we here follow the discussion by Y. Kafri et al., 2002. If we ignore the configurational entropy of a bound segment embedded in the ambient space, its statistical weight will be given by w` = exp(−`E0 /kB T )


where ` is the length of the segment, and E0 is a binding energy in which, for the moment, no distinction is made between the chemical nature of the bases. By contrast, a denaturation loop has no energy associated with it; its statistical weight is consequently determined only by its degeneracy - we have to count the number of its configurations in order to estimate its contribution to the entropy of the configuration.



Assuming that the DNA loop is fully flexible, in the simplest modelling it can be considered as a random walk which returns to its origin after a path of length 2`. From this modelling idea one then knows that the statistical weight for the loop of length ` has an algebraic form, s` (2.7) `c where s is some constant; the amplitude prefactor σ will be simply put to σ = 1 for the moment; but we will come back to it later. Ω(2`) = σ

The exponent c is determined by the statistical properties of the loop configurations. Finally, the configuration of the chain ends must be characterized; they consist of two denatured strands of length ` with a weight Λ(2`) of a similar form as in eq.(2.7), but with a different exponent c. The values c and c can assume will be given later. Given these weights, we can now proceed to calculate the total weight of any given configuration. Supposing the example shown in Figure 2.2, its statistical weight is obtained by Λ(2`0 )w`1 Ω(2`2 )w`3 Λ(2`4 ).


The melting phase transition. This section is somewhat more formal since we now want to determine the nature of the transition between a bound and a denatured phase of a double-stranded DNA molecule. In order to determine the thermodynamic properties of this model, it is practical for the calculation to assume a grand canonical ensemble in which the total chain length L is allowed to fluctuate.2 The reason why this is a good choice is that in this ensemble the total partition function can be conveniently expressed as a geometric series, Z=

∞ X

Z(L)z L =


V0 (z)Q(z) , 1 − U (z)V (z)


where Z(L) is the canonical partition function of a chain of given length L, and z is the fugacity.3 The functions U (z), V (z) and Q(z) are defined as U (z) ≡

∞ X `=1

2 Remember

Ω(2`)z ` =

∞ X (sz)` `=1


= Φc (sz) ,


that we argued in Part I, Chapter 1 that in the thermodynamic limit the ensembles are equivalent. 3 The fugacity is the intensive variable conjugate to the chain length, as much as the chemical potential is conjugate to the particle number.


Computational Biology: A Statistical Mechanics Perspective V (z) ≡

∞ X

w` z ` =


Q(z) ≡ 1 +

∞ X

wz , 1 − wz

Λ(2`)z ` = 1 + Φc (sz)




and V0 (z) = 1 + V (z) .


Φc (sz) is the polylog function which converges for |z| < 1; if in addition Re(c) > 0, the function has an integral representation 1 Φc (z) = Γ(c)



dt t c−1

ze−t , 1 − ze −t


where Γ(c) is Euler’s gamma function. The integral representation allows to see that Φc (z) has a divergence of the form |z − 1|c−1 for z → 1, if c ≤ 1. If c > 1 and 1−z = ε  1, one has Φc (1)−Φc (1−ε) ∼ εζ where ζ = min(1, c−1). In order to fix the average chain length L one has to choose the fugacity such that ∂ ln Z . (2.15) L= ∂ ln z The thermodynamic limit L → ∞ is obtained by allowing z to approach the lowest fugacity value z ∗ for which the partition function Z diverges. This divergence can have two sources: either the numerator grows without bounds, or the denominator vanishes. In fact, both cases occur. At low temperatures, the denominator vanishes, which is at U (z ∗ )V (z ∗ ) = 1 .


Making use of expression eq.(2.11) for V (z), this result can be expressed as U (z ∗ ) = Φc (sz ∗ ) =

1 1 −1= . wz ∗ V (z ∗ )


The solutions of this equation depend on the singularities of Φc (z), which depend itself on the value of c. We take as the order parameter for the denaturation transition in the long chains the fraction of bound monomers4 , θb . Its temperature dependence in 4 For

the case of short chains, θb = θD .



the thermodynamic limit can be calculated from the behaviour of z ∗ (w), since the average number of bound pairs in a chain is given by hmi = so that

∂ ln Z ∂w

hmi ∂ ln z ∗ = . L→∞ L ∂ ln w

θb = lim



In the case of the short chains, we defined the transition as the point where the temperature-derivative of the fraction of bound (or, unbound) base pairs would show a maximum. We also noted that for somewhat longer chains, this curve would steepen, and we interpreted this as a cooperative effect. Indeed, it is a signature of a collective phenomenon, which in the limit of infinite system becomes sharp, provided all bases were to open up collectively. If we now look at the order parameters hmi and θb we defined for the state of the DNA double-strand, we are asked to check what the possible limiting behaviour of these quantities in the limit L → ∞ is. This whole discussion resembles the (simpler) discussion of the phase transition in the 1-dimensional Ising model in Part I of the book. The difference between the two is that, apart from the energy of the configurations, we have to explicitly account for the loop entropy. And that, as we will see, makes everything different since the entropic weights of the loops are algebraic in nature, hence of a long-range nature and not rapidly decaying as for an exponential contribution. As a consequence of this observation, the critical quantity in this discussion is the exponent c. Depending on its value we can find three different scenarios: • c ≤ 1: no phase transition. The function U (z) is finite for all z < 1/s, and it diverges at z = 1/s. The function 1/V (z) is always finite and intersects U (z) for z < 1/s. There is no singular behaviour. • 1 < c ≤ 2: continuous phase transition. In this case, U (z) is finite at z = 1/s since c > 1. For z > 1/s, it is infinite. The singular point zM = 1/s is thus the phase transition point, separating a bound and a denatured regime. At the transition, the derivative of U (z) diverges. • c > 2: first-order phase transition. In this case, U (z) and its derivative are finite at z = zM . Again, there is a transition at this point; above the transition, θb vanishes in the thermodynamic limit. The transition is thus first-order. All that remains to do now is to fix the value of c, which so far has been left unspecified. Since we have modeled the loops as random walks, we have to quantify the relationship between random walks and the polymeric nature


Computational Biology: A Statistical Mechanics Perspective

of DNA. Indeed, if we were to consider the DNA loop simply as a random walk of a given length which returns to the origin, we would find a value of c = d/2 in d space dimensions. Taking this exponent, and looking into the list of possible behaviours, there would be no denaturation transition in d ≤ 2; for 2 < d ≤ 4 the transition is continuous, and for d > 4, the transition will be first-order. This argument can be somewhat refined when considering the loop as a self-avoiding walk, since for a purely random walk configurations can occur in which the walk crosses itself. The assumption of self-avoidance seems more realistic since a real polymer can of course not intersect itself. One has c = dν


where ν is the exponent associated with the radius of gyration5 RG of a selfavoiding random walk. For a walk of length L one has RG ∼ Lν with ν = 3/4 in d = 2 and ν ≈ 0.588 in d = 3. For the loop exponent c, this yields c = 1.5 in d = 2, and c = 1.766 in d = 3. The inclusion of self-avoidance thus leads to a slight smoothing of the transition. All these reasonings are based on the assumption that the loop can be considered an isolated object which does not interact with the rest of the chain. In fact, the whole build-up of the weights assumes that one can consider the ‘bits and pieces’, i.e., the helices and loops, as essentially independent from each other. Recently, however, arguments have been put forward to account for the self-avoidance between the loops and the rest of the chain, albeit in an approximate way. These so-called excluded-volume effects between a loop and the chain have originally been derived using the theory of polymer networks by B. Duplantier, 1986. The discussion of his theory goes beyond the scope of this book since it heavily relies on renormalization group arguments; here, it must suffice us to simply state that the excluded volume interactions modify the self-avoidance exponent relation eq.(2.20) into the expression c = dν − 2ν3


where ν3 takes into account the contribution from the two 3-vertices at the extremities of a loop, which connect the loop to the chain. In d = 3, one finds a value of ν3 ≈ −0.175. Thus for c results c ≈ 2.115


i.e., a value which is slightly larger than two. This result is in accord with independent numerical work based on Monte-Carlo calculations for lattice 5 The

radius of gyration is the average distance of a monomer from the centre of mass of the polymer.



models of DNA (E. Carlon et al., 2002). A side remark on the boundary effects. Just in order to complete the discussion, it turns out that the result on the denaturation transition does not depend on the value of c. This exponent enters the discussion, e.g., if one is interested in the average length of the end segment near the transition, which is given by ∂ ln Q (2.23) ξ=z ∂z z=z∗ with the value of z ∗ determined before. The three different behaviours found for the denaturation transition can be found back for the average end segment length; for more details of this behaviour, see Y. Kafri et al., 2002. Here we simply note that the value of c¯ can be obtained by summing up the scaling dimensions of a linear chain and a fork, leading to c = −(ν1 + ν3 ) ≈ 0.092, i.e., a very small value. Near the melting transition the end segment diverges according to 1 ξ∼ , (2.24) |T − TM | where TM is the melting temperature. This completes this somewhat specialistic discussion. The melting transition with sequence dependence. The main result of the previous section was that the denaturation transition may in fact be a first-order rather than a continuous transition if excluded-volume effects between the bubbles and the chain do matter. This finding has revived a dormant, but long-ongoing controversy on the nature of the transition. We will now turn to the computation of the melting curves of specific DNA sequences and want to see whether the computational results can faithfully represent experimental data. We thus first have to build in sequence-dependence into the theory which we had ignored so far. In order to systematically build in sequence effects we will follow a version of the Poland-Scheraga model which is also useful for further generalizations (like the thermal stability of hybrid DNA, i.e., chains which are not fully complementary and/or of different length). T. Garel and H. Orland, 2004, have recently cast the Poland-Scheraga model into a recursive formulation based on the partition function.6 Previous work by D. Poland, 1974, had relied on a recursion formulation for configuration probabilities. Based on these, the simulation program MELTSIM had been developed by R. Blake 6 We

have seen a basic version of this procedure in the section on the Ising model in Part I of this book, and will encounter another application in the discussion of RNA secondary structure in the following Chapter.


Computational Biology: A Statistical Mechanics Perspective

FIGURE II.2.3: partition function

Graphical representation of the Poland recursion for the

et al., 1999, a variant of which had been used to obtain the simulation results described here. In the original Poland-Scheraga model one only considers two complementary strands of equal length N . Let us call Zf (α) the forward partition function of the two strands starting at base 1, and ending at base α (which is assumed paired). The interaction of the bases is built as that between base pair stacks at positions α and α + 1 along the chain, with associated stacking energies7 α ≡ α,α+1;α,α+1 . (2.25) In order to find the recursion relation obeyed by Zf (α + 1) one has to consider the three possibilities to pair two bases at position α + 1: • the last pair (α, α + 1) is stacked; • there is a loop starting at any α0 with 1 ≤ α0 ≤ α − 1 which ends at α + 1; • there is no loop. This is illustrated in Figure 2.3. Formally, this recursion relation is expressed as Zf (α + 1) = e−βα Zf (α) + σ

α−1 X

Zf (α0 )N (2(α + 1 − α0 )) + ΣM(α) (2.26)

α0 =1

where β = 1/kT . There are two cooperativity parameters in this expression: σ and Σ; these parameters quantify the probabilities for loop and fork formation, respectively, and are assumed to be sequence-independent.

7 The

values for these energy parameters are defined such as to include the hydrogen bond contributions.



The factor N counts the number of conformations of a chain beginning at base α0 and ending at α + 1. Asymptotically, one has 0

N (2(α + 1 − α0 )) = µ α−α f (α − α0 )


where kB ln µ is the entropy per base pair (irrespective of its nature), and f (`) is the return probability of a loop of length 2` to the origin. Finally, in eq.(2.26), M(α) = µα g(α), which counts the number of conformations of a pair of unbound chains starting at base 1 and paired at base α + 1. The function g has a power-law behaviour, however, as we have seen before, with an exponent close to 0. We therefore simply put this factor to 1. In a similar fashion as for the forward partition function Zf one can compute the backward partition function Zb , starting at base N and ending at a paired base α. Again there are three options to pair a base at position α, and thus one finds Zb (α + 1) = e−βα Zb (α + 1) + σ


Zb (α0 )N (2(α0 − α)) + ΣM(N − α) .

α0 =α+2

(2.28) From the expressions eq.(2.26,2.28) one obtains the probability for the binding of a base pair α as Zf (α)Zb (α) , (2.29) pα = Z where Z is the partition function of the two strands, given by  Z = Zf (N ) + Σ µZf (N − 1) + µ2 Zf (N − 2) + ... + µN −1 Zf (1) or, expressed in terms of the backward partition function,  Z = Zb (1) + Σ µZb (2) + ...µN −1 Zb (N ) .



For the implementation of these recursion relations one has to take into account that the algorithm is O(N 2 ) since one has to compute O(α) terms for each value α. In order to reduce the computational complexity, M. Fixman and J. J. Freire, 1977, have developed an approximation in which the powerlaw loop-entropy factor is replaced by a sum of exponentials I

X 1 f (`) = c ≈ ai e−bi ` ` i=1



Computational Biology: A Statistical Mechanics Perspective

where the I parameters (ai , bi ) obey a set of non-linear equations and determine the degree of accuracy. With this step, computational complexity is reduced to O(N · I). In the case of sequences of unequal length, the complexity of the algorithm can reduced with this method from O(N12 N22 ) to O(N1 N2 ). We can turn to the application of the PS-model, and we will see that the story that results will be, to some extent, a story of the cooperativity parameter σ.


The melting profiles of genomic DNA and cDNA

Melting genomic DNA. We now want to take a look at the comparison of the theory we have described with experimental data. In particular we ask what value of c, the one without excluded-volume or with inclusion of excluded-volume effects does fit experiments best? Could we even find out this way whether the transition is first or second order? In order to perform such a quantitative comparison, the theory has finally to be complemented by experimental parameters. The first set of parameters are the nearest and paired neighbor energies (R. D. Blake et al., 1999), distinguishing the different paired bases and their stacks; there are ten independent values to be specified. Further, there is the amplitude factor, the cooperativity parameter σ which we had put to a value of one before. On a technical level, this parameter determines the relative magnitude of energetic and entropic effects. Its name, however, indicates its physical meaning: it determines how many bases interact cooperatively, i.e., open up together to disrupt the helix and form a loop.

FIGURE II.2.4: Schematic diagram of the linearization of an AT-insertion in a GC-rich bacterial plasmid. The bar represents the inserted DNA segment which upon cutting is either placed at an extremity of the molecule (a) or at its center (b). Reprinted with permission from R. Blossey and E. Carlon; Copyright (2003) American Physical Society.



FIGURE II.2.5: Fits of the melting curve with different values of c and σ. The curve with c = 2.15 and σ = 1.26 × 10−5 does not fit the experimental data; by contrast, two choices for c and σ lie exactly on top of each other. Reprinted with permission from R. Blossey and E. Carlon; Copyright (2003) American Physical Society.

A good system to test the issue of the order of the denaturation transition has been introduced by R. D. Blake and S. G. Delcourt, 1998. They inserted artificial AT-rich sequences of varying length into GC-rich bacterial (plasmid) DNA, with sequence lengths between about 200 and 700 bp. After the insertion, the circular plasmid DNA was linearized in two ways: either the insertion was left at the extremity, or left imbedded in the GC-rich chain, see Figure 2.4. Both configurations differ in their melting temperatures: the loop end sequence which melts from the extremity has a lower TM , i.e., TM > TM . Figure 2.5 shows a first example of a calculated differential melting curve for the case of an embedded loop, for the longest inserted sequence with 747 bp, and for three sets of values (σ, c). It is found that the curves with the sets (σ = 1.26 × 10−5 , c = 1.75) and (σ = 1.75 × 10−4 , c = 2.15) fall on top of each other; the curve with the value σ = 1.26 × 10−4 and c = 2.15 has a higher TM . Note that the two curves which fall on top of each other in this graph are in accord with experiment (data not shown). This result shows that a change in the value of c can apparently be compensated for by choosing a smaller value of σ.


Computational Biology: A Statistical Mechanics Perspective

FIGURE II.2.6: Length dependence of the melting temperature shift. Reprinted with permission from R. Blossey and E. Carlon; Copyright (2003) American Physical Society.

Figure 2.6 shows the theoretical values for the melting temperature differloop end ence ∆TM ≡ TM − TM the two sets (σ, c), in comparison with the experimental data, as a function of inverse insert length, 1/N . Both theoretical curves deviate for shorter chain lengths, for which the theory is less reliable (see below). Given that the experimental resolution of melting temperatures is on the order of 0.1o C, there is obviously no direct way to decide the issue of the nature of the transition based on a comparison of the Poland-Scheraga model with this kind of experiment. If one considers longer sequences of genomic DNA, such as, e.g., the human cancer-related gene eIF-4G with a sequence length of about 2900 bp, a structure-rich denaturation profile appears with many interior openings, hence loops, along the sequence. This is illustrated in Figure 2.7, where the differential melting signal is shown together with the denaturation probability 1 − A(i), which is the probability that the i-th base pair is in a closed state. The figure shows this probability at six different temperatures, labelled with greek letters α, .., φ. Again, in this Figure, the melting curves are compared for different values of σ and c, and as before for the plasmid inserts, again the curves can be recovered with two sets of parameter (and hence, consequently, in fact by a whole range of parameter values, interpolating between the two chosen values of σ and c). One may wonder whether the whole discussion misses an important physical parameter of DNA, the stiffness of the double helix, which we encountered in Part I in the discussion of the WLC model. Since the persistence length



FIGURE II.2.7: Melting curve of the cancer-related human gene eIF4G with about 2900 bp (top); opening regions along the chains at different temperatures (bottom). Reprinted with permission from R. Blossey and E. Carlon; Copyright (2003) American Physical Society.


Computational Biology: A Statistical Mechanics Perspective

FIGURE II.2.8: Effect of a change of the cooperativity parameter on the opening probability of a short chain designed for loop opening. Reprinted with permission from R. Blossey and E. Carlon; Copyright (2003) American Physical Society.

of the double-stranded DNA, `ds P , is typically much larger than that of singlestranded DNA, the relevance of excluded-volume effects between the loop and the chain has been questioned (A. Hanke and R. Metzler, 2003). In their view, the double-stranded DNA enclosing the bubble is so stiff compared to the flexible single strands of the bubble that the excluded volume effect plays no role. Several factors, however, intervene in this issue. Firstly, denaturation is a high-temperature phenomenon, since melting temperatures typically lie in the range between 60o C-90 o C, and not at room or physiological temperature, where values of the persistence length are often considered. Secondly, in the limit of high salt concentrations, electrostatics is fully screened and the dependence of `P on temperature can be assumed to follow the worm-like chain model where `P ∼ T −1 (see Part I, Chapter 1). A further effect affecting the value of the persistence length is the presence of small bubbles along the melting chain. It should be noted that the Poland-Scheraga model is tuned to describe long chains and the opening of long loops along the chain; the length of helix and loop segments should scale as the inverse cooperativity parameter, hence as 1/σ. One could thus be misled to assume that short loops would be suppressed during the melting process, which is wrong.



This fact is illustrated in Figure 2.8, which plots the probability of finding a specific base pair, i = 1 and i = 17, in an open state in a chain of 30 bp length. Both A(1) and A(17) can be measured by fluorescence techniques; the base pair i = 17 is placed in the interior of an AT-rich region. The shortness of the sequence leads to the insensitivity of the melting curve on the value of c, but it sensitively depends on σ. For the small value of σ ∼ 10−4 , both probabilities are indistinguishable, which would mean that there is no loop opening, and the sequence rather melts from the ends. Increasing the value of σ by a factor of 100 allows the opening of loops, and brings the theoretical result close to what is observed in experiment (G. Altan-Bonnet et al., 2003). Consequently, it appears that σ = σ(L). We thus conclude that • The Poland-Scheraga can be used to describe even complex denaturation profiles in quantitative accord with experiment; • The theoretical results depend on the value of the critical exponent c and the cooperativity parameter σ, both need to be (and can be) adjusted to fit to experiment; • The fit to experiment depends on the length of the DNA under scrutiny; by construction, the Poland-Scheraga model is more tuned to DNA properties on longer length scales; • We cannot clarify the issue of the order of the phase transition; from a practical point of view for the melting profiles of DNA this seems almost an academic question. The transition is borderline between first and second order, beyond experimental resolution.8 Melting cDNA. DNA contains various sequence regions which are biological entities: first there are the genes, or more specifically, the protein-coding regions, the exons, but also the so-called junk DNA, non-coding regions, the introns. There are also regulatory regions: places where regulatory proteins attached and control the expression of genes (see Part III). If we melt purely genomic DNA, we will get a bit of everything. Can we distinguish between the melting behaviour of different sequence regions - is there a correlation between the biological purpose of a sequence and its (thermal) stability ?9 8 Physicists

protesting against this pessimistic conclusion should look back into the critical properties of classical superconductors. 9 One might object that the thermal stability of DNA cannot be relevant for organisms since the melting temperatures of DNA are usually much higher than physiologically relevant temperatures. One counter argument against this is that a position which is easier to destabilize thermally can likewise be more readily destabilized either by changes of chemical conditions (pH) or the localized action of forces, e.g., by protein complexes on DNA.


Computational Biology: A Statistical Mechanics Perspective

We address this question for complementary DNA (cDNA). Complementary DNA is a molecule which contains only exons (including so-called untranslated end regions, UTR’s). Complementary DNA can be obtained from genomic DNA by first transcribing it into RNA, splicing out of the introns, and transcribing the mature RNA back into DNA; this latter step is done by a viral enzyme, the reverse transcriptase. Figure 2.9 shows the build-up of genomic DNA and cDNA in a schematic comparison.

FIGURE II.2.9: Structure of genomic DNA vs. cDNA. Reprinted with permission from E. Carlon et al.; Copyright (2005) American Physical Society.

Melting of genomic DNA of various organisms has been studied by E. Yeramian, 2000. Quite generally it is found that the melting curves reflect the base composition; since exons are on average more GC-rich than introns, which have a bias to AT bases, this difference can in some cases permit to distinguish genes and non-coding regions from each other. AT-rich regions melt more easily since the base pairing has only two hydrogen bonds, instead of the three between G and C. However, this distinction does not always work; e.g., for bacteria it was found that the structure of their ultra-dense genomes without non-coding regions does not permit to identify genes based on melting signatures alone. This is indeed different for cDNA (E. Carlon et al., 2005). If one melts cDNA, one finds a differential melting curve similar to the one shown in Figure 2.10. As for genomic DNA, several distinct melting peaks arise. In the figure, sequence information has been added: the horizontal bars indicate



the location of the boundaries between two exons as known from sequence annotation. The numbers on the right indicate the number of base pairs of the intron removed at that position. The vertical bars in the figure indicate when the probability of having bound pairs falls below 1/2. The ends of these bars locate the position of a thermal boundary: a sequence position at which the opening of the sequence is stopped at increasing temperature. Figure 2.11 shows that in several cases thermal boundaries coincide with the annotated exon-exon boundaries, i.e., known coding regions of the genes. This observation holds for many human genes, with a resulting coincidence of thermodynamic and annotated boundaries at about 30 %. At present, there is no simple explanation for this finding; but one may speculate. In this context it is useful to think in evolutionary terms. How did exons and introns come into their positions in the first place? There are two current, opposing hypotheses on the evolution of genomes, one based on the idea that exon positioning came late, and were inserted (shuffled around) into intronic DNA. (Bacteria have no introns: but according to the theory they were just very active in getting rid of the junk.) The opposing hypothesis considers exons to be evolutionarily ‘early’, and introns being inserted into an otherwise largely exonic DNA (W. Gilbert, 1987). The finding that exon-exon-boundaries are, in humans, to about 30 % located in positions at which DNA is less (thermally) stable might support the view that introns are ‘late’: the location of thermal boundaries might be preferred sites for intron insertion, since in these positions double-stranded DNA is more readily opened up. But then, these spots might also be prone to intron losses. In any case, an agreement between positions at which a physical signature coincides with a biological signature to such an extent seems hard to be just accidental.


Computational Biology: A Statistical Mechanics Perspective

FIGURE II.2.10: Melting curve and exon-exon boundaries. For the explanation, see text; N is sequence length. Reprinted with permission from E. Carlon et al.; Copyright (2005) American Physical Society.

FIGURE II.2.11: Correlation of thermal boundaries with exon-exon boundaries. Reprinted with permission from E. Carlon et al.; Copyright (2005) American Physical Society.




Hybridizing DNA

The final section of the Chapter on DNA melting is devoted to a discussion of the hybridization of oligomeric DNA on microarrays in order to measure gene expression levels. The idea of this method is that the amount of RNA in a cell (or, of back-transcribed RNA into cDNA) can serve as a measure for the number of proteins that are translated in the cell; this is only approximately correct, but nevertheless an indication. The recognition capability of a single-stranded RNA or DNA towards its complementary sequence can be exploited in a very simple way to measure the presence and amount of DNA in a given sample. In order to put this idea into practice, it has now become possible, using techniques from the microelectronics industry, to produce biochips - in analogy to microelectronic chips. The principle of a biochip, or DNA microarray in our context, is the following. On a substrate, a collection of single-stranded DNA molecules, the probes, is built. (We will not bother here with the technical details how this is done; the interested reader is asked to consult the literature indicated at the end of the Chapter.) In the case of Affymetrix chips, the only type of microarrays we address here, the sequences contain 25 base pairs which have been selected from a gene of interest. In fact, about 10-16 probes are produced and fixed at the substrate, reflecting different selected subsequences from the same gene. In addition, for each perfectly matching probe - one that is exactly identical to a subsequence from the selected gene - a mismatching probe is produced. This probe differs from the perfect match by one nucleotide. The mismatch probe is a control, which is used to assess cross-hybridizations, i.e. non-specific hybridizations. Based on this idea microarrays containing full genomes can now be built. They are used by exposing the probe strands to a set of target molecules which will bind to the selected probes, as schematically shown in Figure 2.12. The amount of bound DNA is measured, typically optically by using fluorescently labelled target RNA. How do targets and probe meet each other on the chip? A very simple one is the following (E. Carlon and T. Heim, 2005), motivated by previous work by G. A. Held et al., 2003. After hybridizing targets and probes light intensities I of the fluorescent markers are measured. Distinguishing between specific, S, and non-specific hybridizations, N , we write I(c, ∆G) = S(c, ∆G) + N + 



Computational Biology: A Statistical Mechanics Perspective

FIGURE II.2.12: Hybridization on a microarray. Note that in the case of Affymetrix arrays, probes are DNA, while targets are RNA. where  stands for purely experimental noise. I is the intensity from the probe whose complementary target RNA is present in solution at a concentration c, and ∆G is the hybridization free energy - which we had discussed in Section 1.1. A distinction between mismatching or matching DNA does not need to be made, since both will differ only in their values of ∆G. The non-specific hybridization contribution N depends on total RNA concentration in solution, and probably other free energy parameters reflecting partial hybridizations. The value of N is not important if one decides to consider the quantity ∆I ≡ I(c) − I(0) ≈ S(c, ∆G)


Such a background subtraction is possible when we compare measurements in which c is the concentration of a particular gene; this is the case for the set of control measurements done in a Latin square, which consists of a well-defined dilution series for a set of specific genes added to a background, the so-called ‘spike-in’ genes, and, importantly, contains the case c = 0. In the simplest case, which is well applicable to chains of the order of 25 base pairs as we know from Section 2.1 of this Chapter, we can understand the binding of the probe DNA and the target RNA as a two-state process, in which the target is either fully unbound in solution, or fully bound at the correct sequence at the surface. The binding is then described by a Langmuir isotherm of the form Ace−β∆G (2.35) S(c, ∆G) = 1 + ce−β∆G where, as before, β = 1/RT . The amplitude factor A sets an intensity scale;



FIGURE II.2.13: Hybridization data falling on the Langmuir isotherm; plotted against the two scaling variables x (top) and x0 (bottom) defined in the text. Reprinted with permission from E. Carlon and T. Heim; Copyright (2005) by Elsevier.

it corresponds to the saturation value in the limit c  eβ∆G , i.e., whenever the concentration is high or the binding energy large. Figure 2.13 shows an example of such data analysis. Noticing from eq.(2.35) that the free energy and the concentration appear always in the form of a product x ≡ c exp(−β∆G), we can use x as a scaling variable. The data then fall nicely on the Langmuir isotherms. In these curves, the ∆G values were computed with parameters determined experimentally for DNA-RNA hybrids in solution and not for the DNA attached to the surface, and we may assume a systematic deviation between the


Computational Biology: A Statistical Mechanics Perspective

FIGURE II.2.14: Schematic drawing of RNA hybridization in solution indicating two possible effects, self-hybridization of targets and hybridization of different fragments. (known) solution values and the (unknown) surface values. This systematic deviation between the free energy values in solution and at the surface can likewise be shifted into one overall parameter: temperature. Temperature T then changes status from an experimental to a fitting parameter. If its value is chosen to lie at around 700 K, i.e., roughly twice as large as the true experimental temperature, a very good agreement between experiment and the simple theory is reached. As seen in Figure 2.13, only the data from probe 16 deviate significantly from the computed Langmuir isotherm, but they seem to follow another isotherm which appears shifted to the right. This behaviour can be captured by introducing a probe-dependent parameter αk = (1 + c˜ exp(−β∆GRN A )) −1 ,


which takes into account the effect of RNA-RNA hybridization in solution, leading to secondary structures of the RNA strands closing in on themselves, thereby reducing the amount of available sequences for binding at the surface. This effect can be taken into account in the parameter combination c˜ exp(−β∆GRN A ), wherein c˜ is a fitting parameter. Figure 2.13 displays the full data collapse if the scaling variable x0 = αk c exp −β∆G


is used. What happens during hybridization in solution is indicated schematically in Figure 2.14. Sequences with high free energies tend to hybridize to



other chains present in the solution, and they are thus not available anymore for the hybridization process at the surface - their ‘active’ concentration is reduced as compared to that of other sequences. Thus, at the end of the chapter on DNA, we see how relevant the formation of secondary structures can be even for DNA-RNA microarrays. The conclusion is that we need to learn more about this process, and we are thus ready to leave DNA for the RNA world.


Computational Biology: A Statistical Mechanics Perspective

Additional Notes Short DNA’s remain of interest due to their applicability in biotechnology, in particular in microarrays. Modifications of the linear DNA have become of interest as well; the simplest such systems are DNA hairpins, in which a single-stranded DNA bends back to itself and increases specificity, the socalled molecular beacons (G. Bonnet et al., 1999). Short chains can also be used to decorate nanoparticles which enable to establish a link between molecular recognition and phase behaviour (D. B. Lukatsky and D. Frenkel, 2004). Finite-size scaling. Real DNA sequences always have a finite length L, thus one may wonder whether the whole discussion of phase transition effects is at all applicable. There is a systematic way to extract the effect of finite system size at phase transitions which runs under the name finite-size scaling; for the Poland-Scheraga model, this has only recently been discussed ¨fer, 2005). The conclusions are not much different from what was (L. Scha discussed here, though. Other models used in studies of DNA denaturation. There are several other models that have been developed for the description of DNA denaturation aside from the Poland-Scheraga model. We here comment only on one model which follows a rather different philosophy: the DauxoisPeyrard-Bishop model by Th. Dauxois et al., 1993, following earlier work by M. Peyrard and A. R. Bishop, 1989. Their model is more microscopic than the Poland-Scheraga model; it starts out from the dynamics of the chains, i.e., the main ingredient in the model is the transverse stretching yn of the hydrogen bonds between the complementary bases counted by the index n. The model is defined by the Hamiltonian  X 1 2 my˙ + V (yn ) (2.38) H= 2 n n where m is the mass of the bases in the kinetic energy. The potential model contains two contributions    2 k  2 (yn +yn−1 ) −αn yn 1 + %e (yn − yn−1 ) . (2.39) V (yn ) = Dn e −1 + 2 The Morse potential in the first term describes the effective interactions between complementary bases: it contains both the attraction due to the hydrogen bonds forming the base pairs and the repulsion of the negatively charged phosphates in the backbone of the strands, which is screened by the surrounding solvent. The parameters Dn and αn distinguish between the two complementary base pair combinations at site n, and hence induce a sequence dependence. The second term comprises the stacking interactions. The exponential term modifies an otherwise harmonic potential. This nonlinearity



FIGURE II.2.15: Unzipping and thermal denaturation phase diagram

is essential: representing local constraints in nucleotide motions, it induces long-range cooperative effects. The stiffening of the coupling in the bound state compared to that in the open state leads to an abrupt entropy-driven transition. Similar conclusions were drawn by D. Cule and T. Hwa, 1997. The Dauxois-Peyrard-Bishop-model has, as the Poland-Scheraga model, also been confronted with experimental data, in particular for short sequences (A. Campa and A. Giansanti, 1998). Moreover, it has recently been used to correlate the dynamics of loop openings in the chain with regulatory sequence signatures. Correlations of regions with a high probability of loop openings with transcription start sites have been reported (C. H. Choi et al. 2004 and G. Kalosakas et al., 2004). More recent results are more skeptical, see T. S. van Erp et al., 2005. DNA unzipping. The opening of DNA does not necessarily have to arise by thermal denaturation - which typically occurs at non-physiological temperatures, and hence cannot play any direct role in a living cell. Other factors, such as a change of pH, or the direct application of a local force have, however, a similar effect on the molecule. As we have seen already in Part I, pulling experiments on DNA have become a standard approach due to the development of new methods of micromanipulation. From the point of view of the denaturation phase transition, the opening of the chains under the action of a force has been termed ‘unzipping transition’ (for obvious reasons). It can be shown to be a first-order transition, as much as the phase transition in the


Computational Biology: A Statistical Mechanics Perspective

Ising model in a magnetic field, see Part I, Chapter 1. Theoretically, the unzipping transition can be described by a variant of the Worm-Like Chain model, in which in addition to the bending energy and the applied force a binding potential between the two chains is taken into account (for a brief review, see S. Cocco et al., 2002). From the calculation of the free energy of this model one finds a phase diagram in the temperature-force plane, shown in Figure 2.15, which qualitatively indicates the region in which the transitions occur. This figure is only of limited generality: the precise location of the first-order unzipping line can be anywhere inside the range set by the dashed lines, since it is in general sequence-dependent. DNA microarrays. Despite the enormous interest in the use of DNA microarrays, many of their basic physical properties remain fairly poorly understood; an overview of hybridization protocols before the arrival of the high-throughput techniques is by J. G. Wetmur, 1991. The mechanism of probe-target encounter relies on the efficient diffusion of the target molecules on the surface; kinetic modelling efforts have been devoted to model the diffusion problem, see D. Erickson et al., 2003. It is of utmost importance that the fluorescence measurements are done when thermal equilibrium has been reached, as pointed out by G. Bhanot et al., 2003. Conceptually basic modelling approaches based on the Langmuir isotherm approach are reviewed by A. Halperin et al, 2006. The paper that inspired the approach discussed here is by G. A. Held et al., 2003. It should be compared with papers from a more statistical approach, as e.g. L. Zhang et al., 2003. Physics-based approaches tend to be based on few assumptions, and a small number of parameters, quite in contrast to purely statistical approaches.



References G. Altan-Bonnet, A. Libchaber and O. Krichevsky, Bubble Dynamics in Double-Stranded DNA, Phys. Rev. Lett. 90, 138101 (2003) J. Bayer, J. O. R¨ adler and R. Blossey, Chains, Dimers, and Sandwiches: Melting Behaviour of DNA Nanoassemblies, Nano Letters 5, 497-501 (2005) G. Bhanot, Y. Louzoun, J. Zhu and C. DeLisi, The Importance of Thermodynamic Equilibrium for High Throughput Gene Expression Arrays, Biophys. J. 84, 124-135 (2003) R. D. Blake, Cooperative Lengths of DNA During Melting, Biopolymers 26, 1063-1074 (1987) R. D. Blake and S. G. Delcourt, Thermal Stability of DNA, Nucl. Acids Res. 26, 3323-3332 (1998) R. D. Blake, J. W. Bizarro, J. D. Blake, G. R. Day, S. G. Delcourt, J. Knowles, K. A. Marx and J. Santa Lucia, Jr, Statistical mechanical simulation of polymeric DNA melting with MELTSIM, Bioinformatics 15, 370-375 (1999) R. Blossey and E. Carlon, Reparametrizing the loop entropy weights: Effect on DNA melting curves, Phys. Rev. E 68, 061911 (2003) G. Bonnet, S. Tyagi, A. Libchaber and F. R. Kramer, Thermodynamic basis of the enhanced specificity of structured DNA probes, Proc. Natl. Acad. Sci. USA 96, 6171-6176 (1999) A. Campa and A. Giansanti, Experimental tests of the Peyrard-Bishop model applied to the melting of very short DNA chains, Phys. Rev. E 58, 35853588 (1998) E. Carlon, E. Orlandini and A. L. Stella, Role of Stiffness and Excluded Volume in DNA Denaturation, Phys. Rev. Lett. 88, 198101 (2002) E. Carlon, M. L. Malki and R. Blossey, Exons, introns and DNA thermodynamics, Phys. Rev. Lett. 94, 178101 (2005) E. Carlon and T. Heim, Thermodynamics of RNA/DNA hybridization in high density oligonucleotide microarrays, Physica A 362, 433-449 (2006) C. H. Choi, G. Kalosakas, K. O. Rasmussen, M. Hiromura, A. R. Bishop and A. Usheva, DNA dynamically directs its own transcription initiation, Nu-


Computational Biology: A Statistical Mechanics Perspective

cl. Acids Res. 32, 1584-1590 (2004) S. Cocco, J. F. Marko and R. Monasson, Theoretical models for singlemolecule DNA and RNA experiments: from elasticity to unzipping, C. R. Physique 3, 569-584 (2002) D. Cule and T. Hwa, Denaturation of Heterogeneous DNA, Phys. Rev. Lett. 79, 2375-2378 (1997) T. Dauxois, M. Peyrard and A. R. Bishop, Entropy-driven DNA denaturation, Phys. Rev. E 47, R44-R47 (1993) D. Erickson, D. Li, and U. J. Krull, Modeling of DNA hybridization kinetics for spatially resolved biochips, Anal. Biochem. 317, 186-200 (2003) T. S. van Erp, S. Cuesta-Lopez, J.-G. Hagmann and M. Peyrard, Can One Predict DNA Transcription Start Sites by Studying Bubbles?, Phys. Rev. Lett. 95, 218104 (2005) M. Fixman and J. J. Freire, Theory of DNA Melting Curves, Biopolymers 16, 2693-2704 (1977) Th. Garel and H. Orland, Generalized Poland-Scheraga model for DNA hybridization, Biopolymers 75, 453-467 (2005) W. Gilbert, The Exon Theory of Genes, vol. LII Cold Spring Harbour Symposia on Quantitative Biology (1987) A. Halperin, A. Buhot and E. B. Zhulina, On the hybridization isotherms of DNA microarrays: the Langmuir model and its extension, J. Phys.: Condens. Matter, to appear (2006) A. Hanke and R. Metzler, Comment on: “Why is the DNA denaturation transition first order?”, Phys. Rev. Lett. 90, 159801 (2003) G. A. Held, G. Grinstein and Y. Tu, Modeling of DNA microarray data by using physical properties of hybridization, Proc. Natl. Acad. Sci. USA 24, 7575-7580 (2003) Y. Kafri, D. Mukamel and L. Peliti, Melting and unzipping of DNA, Eur. Phys. J. B. 27, 135-146 (2002) G. Kalosakas, K.O. Rasmussen, A. R. Bishop, C. H. Choi and A. Usheva, Sequence-specific thermal fluctuations identify start sites for DNA transcrip-



tion, Europhys. Lett. 68, 127-133 (2004) D. B. Lukatsky and D. Frenkel, Phase Behavior and Selectivity of DNALinked Nanoparticle Assemblies, Phys. Rev. Lett. 92, 068302 (2004) R. Owczarzy, P. M. Vallone, F. J. Gallo, T. M. Paner, M. J. Lane, A. S. Benight, Predicting Sequence-Dependent Melting Stability of Short Duplex DNA Oligomers, Biopolymers 44, 217-239 (1998) M. Peyrard and A. R. Bishop, Statistical Mechanics of a Nonlinear Model for DNA Denaturation, Phys. Rev. Lett. 62, 2755-2759 (1989) D. Poland, Recursion Relation Generation of Probability Profiles for SpecificSequence Macromolecules with Long-Range Correlations, Biopolymers 13, 18591871 (1974) D. Poland and H. A. Scheraga, Theory of helix-coil transitions in biopolymers, Academic Press (1970) J. Santa-Lucia Jr., A unified view of polymer dumbbell, and oligonucleotide DNA nearest-neighbor thermodynamics, Proc. Natl. Acad. Sci. USA 95, 1460-1465 (1998) L. Sch¨ afer, Can Finite Size Effects in the Poland-Scheraga Model Explain Simulations of a Simple Model for DNA Denaturation, preprint cond-mat/ 0502668 (2005) R. Thomas, Recherches sur la d´enaturation des acides desoxyribonucl´eiques, Biochim. Biophys. Acta 14, 231-240 (1954) R. M. Wartell and A. S. Benight, Thermal denaturation of DNA molecules: a comparison of theory with experiment, Phys. Rep. 126, 67-107 (1985) J. G. Wetmur, DNA Probes: Applications of the Principles of Nucleic Acid Hybridization, Crit. Rev. Biochem. Mol. Biol. 26, 227-259 (1991) E. Yeramian, Genes and the physics of the DNA double-helix, Gene 255, 139-150 (2000) E. Yeramian, The physics of DNA and the annotation of Plasmodium falciparum, Gene 255, 151-168 (2000) E. Yeramian, S. Bonnefoy and G. Langsley, Physics-based gene identification: proof of concept for Plasmodium falciparum, Bioinformatics 18, 190-193


Computational Biology: A Statistical Mechanics Perspective

(2002) L. Zhang, M. F. Miles and K. D. Aldape, A model of molecular interactions on short oligonucleotide microarrays, Nat. Biotech. 21, 821-828 (2003); Erratum ibid, 21, 841 (2003).

Chapter 3 RNA


Computing RNA secondary structure: combinatorics

We have seen in the previous chapters why the knowledge of RNA secondary structure is important - for information processing in biology and biotechnology. In this Chapter we want to learn how to predict RNA structure. Predicting RNA structure, as we will see, has three aspects. The first aspect is combinatorial: how can we classify and compute the possible structures? The second aspects relates to the question which of these structures are selected based on their free energy: from the set of all possible structures, which is the one that is energetically the most favorable one? The third and final aspect touches upon the kinetics of the fold formation: how does the molecule actually acquire its fold? Maximal base pairing. We begin the discussion by formalizing the graphical representations introduced in Chapter 1. If the primary structure of an RNA molecule - i.e., its sequence - of length n is denoted by a string r = r1 ...rn , (3.1) its secondary structure can be described as a set S of disjoint pairs (ri , rj ) for 1 ≤ i < j ≤ n. Considering the bases as vertices on a graph, and all pairings as the edges of that graph, the secondary structure is a matching in a graph G = (V, E) with the properties that V contains a vertex for every base pair ri , i = 1, ..., n, and E contains an edge (u, v) if and only if u, v  V are complementary bases. In the prediction of secondary structure, we want to first exclude pseudoknots. In the above notation such a knot occurs when a base ri is paired with a base rj , and a base rk with rl such that i < k < j < l: the pairs overlap. It is this condition which we reject for the moment. The prescription we have given allows to cover all possible secondary structures, except for the occurrence of pseudoknots. However, we would not know



Computational Biology: A Statistical Mechanics Perspective

FIGURE II.3.1: Dynamic programming recursion for RNA structure (see text).

which one to select from the set of structures. A natural selection principle is based on free energy. Assuming that we can associate with every complementary base pair an energy contribution α(ri , rj ) =P−1 and put α = 0 otherwise, we see that the minimal (free) energy E(S) = (ri ,rj ) α(ri , rj ) will be attained for the RNA structure with the maximal number of base pairs. In this simplified version which treats all bound pairs on equal footing, the problem of RNA secondary structure determination can be solved by a dynamic programming approach. With the assumption that α(ri , rj ) is independent of all other pairs and positions of ri and rj in the sequence, we can compute the free energy of a substring ri ...rj of the RNA molecule disregarding the surrounding r1 ...ri−1 and rj + 1...rn . In this way, we can use solutions for smaller strings to obtain those for larger strings and obtain a recursive solution. The resulting dynamic programming algorithm is easy to develop; the possibilities that can arise in building up the structure are shown in Figure 3.1. There are three alternatives: • Add paired bases to an optimal structure for the subsequence i+1, j −1; • Add an unpaired base to an optimal structure for the subsequence i, j−1; this possibility arises symmetrically for the other end of the molecule; • Combine two optimal substructures i, k − 1 and k, j. The latter step obviously runs into problems when pseudoknots are to be accounted for.



The resulting recursion can be summarized as  E(Si+1,j−1 ) − 1          E(Si+1,j ) E(Si,j ) = min    E(Si,j−1 )       min{E(Si,k−1 ) + E(Sk,j )} , i < k ≤ j


In order to initialize the algorithm, one requires E(Si,i ) = 0 for i = 1, .., n and E(Si,i−1 ) = 0 for i = 2, .., n. Let us illustrate this scheme with a small example (S. R. Eddy, 2004). We take the RNA sequence r = GGGAAAU CC; for each allowed pair AU , GC we take α = −1, and α = 0 otherwise. The pairing matrix (r × r) at the initialisation step reads   GGGAAAU CC  G 0    G 0 0    G x 0 0    A x x 0 0   (3.3)  A x x x 0 0    A x x x x 0 0    U x x x x x 0 0    C x x x x x x 0 0 Cx x x x x x x 0 0 where the lower half (x) is not used, and the diagonal and its left neighbour are put to 0. In the computation, one finds that entry Y depends on all entries denoted by y, and w = (i + 1, j),  G  G  G  A  A  A  U  C C

G 0 0 x x x x x x x

GGAAAU CC 0 0 x x x x x x

0 0 x x x x x

y 0 0 x x x x

y y y w 0 0 0 x 0 0 x x 0 x x x

Y y y y y 0 0 0

               



Computational Biology: A Statistical Mechanics Perspective

FIGURE II.3.2: gorithm.

Exemplary RNA structure obtained from the simple al-

so that one obtains the following result by going through the recursion   GGGAA A U C C  G 0 0 0 0 0 0 −1 −2 (−3)     G 0 0 0 0 0 0 −1 −2 (−3)     G x 0 0 0 0 0 −1 (−2) −2     A x x 0 0 0 0 (−1) −1 −1    (3.5)  A x x x 0 0 (0) −1 −1 −1     A x x x x 0 (0) −1 −1 −1    U x x x x x 0 0 0 0    C x x x x x x 0 0 0  C x x x x x x x 0 0 The minimum structure can be found by tracing back through the table following a diagonal trace (i, j) to (i + 1, j − 1). The result is shown in Figure 3.2; the trace back is indicated by terms in brackets. The complexity of the algorithm to compute the structure is O(n3 ), since there are n2 entries, and each computation is O(n). The trace back to find the structure takes again linear time O(n), if back pointers are used. Following the original approach by R. Nussinov and A. B. Jacobson, 1980, there have been several further developments of algorithms for RNA secondary structure prediction. M. Zuker, 1989, made an important contribution by developing a tool for the determination of sub-optimal folds. More recent developments are targeted towards the inclusion of pseudoknots; we comment on them in the Additional Notes. The basic approach to RNA secondary structure, as we have seen, is to compute the possible combinations by a recursion, and to score them by a suitably defined free energy. Thinking back to what we saw for DNA, one might question whether the first part - the computation of the configurations - cannot also be done based on our main tool from statistical mechanics, the



partition function Z. This is what follows.


The RNA partition function

The partition function of a chain of L nucleotides can be written in the following very general way as (H. Orland and A. Zee, 2002) Z=


d3 rk F({r}) ZL ({r}) .



Here, rk is the position vector of the k-th nucleotide. The function F({r}) =

L−1 Y

f (|ri+1 − ri |)



is a model-dependent function of the molecular geometry, and takes into account the steric constraints of the chain. Standard choices for f are f (r) = δ(r − `) ,


if the nucleotides are connected by rigid rods of size `, or f (r) = exp[−(r − `)2 /6η 2 ]


if the rods are taken as elastic springs with η as a measure of their stiffness. The partial partition function ZL in eq.(3.6) counts the different configurations of paired bases. It can be defined by the series X X ZL ({r}) = 1 + Vij (rij ) + (Vij (rij )Vkl (rkl ) + Vik (rik )Vjl (rjl )) + ... (ij)


(3.10) with rij ≡ |ri − rj |, and where the summation index (ij) denotes all pairs with i < j, (ijkl) all quadruplets with i < j < k < l, and so forth. The first term of the series describes the binding energy between the bonds i and j, with the indices running all along the chain, the second the binding of i and j in a configuration together with k and l, and so forth. The factors Vij ≡ exp(−βεij vij (rij ))θ(|i − j| > 4)



Computational Biology: A Statistical Mechanics Perspective

are the Boltzmann factors associated with a (4 × 4)-dimensional symmetric matrix εij of bond energies between the i-th and j-th bases at a distance rij ; β = 1/(kB T ) is the inverse thermal energy as before. In eq.(3.11), the factor vij = v(|ri − rj |) is a short-range attractive interaction between the bases. The Heaviside function θ(|i − j| > 4) expresses the steric constraint which prohibits hybridization of the bases in closest proximity to each other. Finally, note that Vii = 0. The series (3.10) can be expressed in integral form by introducing a set of i = 1, ..., L Hermitian matrices1 ϕi , i = 1, ..., L, using the expression for the ordered matrix product Oπ [{ϕ}] ≡

L Y (1 + ϕl ) ≡ (1 + ϕ1 )(1 + ϕ2 ) · · · (1 + ϕL ) .



The matrices have the dimension (N × N ). The parameter N is introduced on purely formal grounds; as will be seen below, it allows to organize the expansion of the partition function according to the topology of RNA. It can, however, also be related to physical quantity, the concentration of a chemical which favors pseudoknots. Such chemicals, e.g., are divalent ions such as Mg2+ , and the relationship is given via the idenfication N −2 ≡ exp βµ where µ is the chemical potential of the ions. Coming back to the calculation, the resulting formula for ZL (N ) is ZL (N ) = AL (N )




dϕk e− 2


ij (V


)ij T r(ϕi ϕj )


1 T rOπ [{ϕ}] N


where the normalization factor is given by AL (N ) =



dϕk e− 2


ij (V


)ij T r(ϕi ϕj )




In both expressions, V is an (L × L)-matrix with entries Vij . After the introduction of the matrices, the partition function given by eq. (3.13) looks like a Gaussian integral in the matrices ϕk over a product observable in the ϕk . The product of the terms (1 + ϕl ) evaluates into a polynomial of order L and hence we have to perform Gaussian integrals over all the contributions of this polynomial from order 1 to L - thus nothing but 1A

square matrix is hermitian if it is self-adjoint, i.e., fulfills Aij = Aji , where the overbar stands for the complex conjugate. We encountered examples of self-adjoint matrices in Part I, Chapter 1, in the discussion of the Ising model, e.g., the matrix σ3 .



the moments of the partition function. The important thing is now that the introduction of the matrix-dimension N gives a handle to reorganize the series representation. One can show that the result is an asymptotic series in 1/N of the form ZL (N ) = 1 +

X i R it is transformed into d2 ϕ + κ2 e −ϕ(x)+2x/R = 0 . (4.43) dx2 The shifted potential ϕ˜ = ϕ(x)−2(x/R) satisfies the planar Poisson-Boltzmann equation, albeit with a different boundary condition dϕ(x) ˜ `B σ 2 = − . (4.44) dx x=0 Ze R The solution of the Poisson-Boltzmann equation is, by analogy to the planar case, given by   κx , (4.45) ϕ˜ = 2 ln 1 + √ 2


Computational Biology: A Statistical Mechanics Perspective

where the boundary condition fixes the value of   1 `B σ 2 κ= − . 2 Ze R


This result makes no sense for `B Rσ < 2Ze, i.e. κ < 0. We then have to take the solution of the Poisson-Boltzmann equation for κ = 0, and obtain instead a logarithmic potential ϕ(r) = 2ξm ln(r/R) (4.47) where ξm ≡

`B Rσ 2Ze


is the Manning parameter. We thus obtain the full solution as  2ξm ln(r/R) , ξm ≤ 1 , ϕ(r) = 2 ln(r/R) + 2 ln[1 + (ξm − 1) ln(r/R)] , ξm > 1 .


For ξm > 1, the electrostatic potential behaves like ϕ(r) ∼ 2 ln(r/R), and is essentially independent of charge density; the counterions are bound to the cylinder. For ξm ≤ 1, the number of counterions cs (r) ∼ exp − ϕ(r) contained in a cylindrical shell R0 around the cylinder behaves as R0 cs (r) ∼ r2(1−ξm ) (4.50) R

and clearly diverges with R0 : the counterions escape to infinity. The phenomenon of the counterion confinement to the DNA is called Manning condensation (G. S. Manning, 1969). Protein electrostatics. We will now go one step further and apply our knowledge to proteins, in particular to the determination of solvation free energies of atoms, ions and biomolecules. In many cases, the dominating contribution is due to electrostatics (B. Honig and A. Nichols, 1995), and the main task then is to compute the electrostatic potential of the proteins and their complexes. The electrostatics problem associated with proteins can be determined by solving the Poisson or Poisson-Boltzmann equation for the geometry depicted in Figure 4.2. It is assumed that the space Ω is filled with an arrangement of charges (not shown explicitly), while we ignore the distribution of ions that can usually be assumed to surround the protein. We thus ‘only’ have to solve the Poisson equation in this inhomogeneous space. The inclusion of the mobile charges is, of course, possible.



FIGURE II.4.2: Electrostatic problem of a protein in a solvent: a body Ω imbedded in space Σ; Γ defines the boundary. Reprinted with permission from A. Hildebrandt et al.; Copyright (2004) American Physical Society. The mathematical formulation of the problem consequently is:

∆φΩ (r) = −

4π % Ω


∇r [Σ (r)∇r φΣ (r)] = 0


∂n [Ω φΩ − Σ φΣ ]|Γ = 0


φΩ |Γ = φΣ |Γ


For proteins, Ω ≈ 2 − 5 is usually assumed, while for the surrounding water, Σ ≈ 80. In the Poisson equation for the solvent we have allowed for a (local) spatial dependence of Σ ; this is frequently done in a phenomenological way to mimic effects of the modification of the water dielectric properties near the protein. This effect is based on the restricted motion of the water near the protein, bringing about distinct orientational polarization correlations. We will see how this effect of water structure can be built into the theory of electrostatics in a more fundamental way in the next subsection. The solution of the above equations can be obtained only numerically for the complex protein surfaces. Several sophisticated solvers exist for this purpose meanwhile; one of the most advanced examples come from the Holst group which was used to calculate the electrostatic potentials of large biomolecular


Computational Biology: A Statistical Mechanics Perspective

structures, such as microtubules and ribosomes (N. Baker et al., 2001), see Color Figure 4. Nonlocal electrostatics. If we want to describe the electrostatic properties of biomolecules on length scales for which a membrane can be modelled as a fluctuating sheet, a DNA molecule as a cylinder, and a globular protein as a sphere, the level of description of electrostatic interactions is well justified. For the large protein structures we have just seen, which are cylindrical on a mesoscopic scale, but have a lot of atomistic structural detail which may - or will - be of biological relevance, we have been pushing things a bit too far. It seems hard to justify why the water surrounding a protein surface can, on this atomic scale, be considered a dielectric medium of dielectric constant  = 80. On such length scales, the usual continuum approach ultimately must break down, since water is not featureless on these scales: water has structure. The water molecules respond to the presence of charges, and their network of hydrogen bond has to rearrange, leading to correlations of the water molecule orientations over a characteristic length scale which we denote as λ. One way to ‘repair’ the error made in classical electrostatics is to modify the dielectric function. Conventionally this is done by an inclusion of a local spatial dependence (r) near the protein surface. This function, which can be parametrized in various ways, is used to effectively reduce the dielectric constant of water to a small value near the protein surface. This approach is, while physically justifiable, a technically uncontrolled procedure. But there is, at least at present, no real practical alternative to continuum electrostatics as it still is much more computationally efficient than microscopic simulations based on, e.g., molecular dynamics (MD) (T. Simonson, 2001). Interest has therefore risen in systematic extensions of the theory of continuum electrostatics that allow to account for spatial variations of the dielectric behaviour of the solvent, in particular near the boundary of a protein. Within the continuum theory of the electrodynamics of matter, the orientational polarizability of water around a protein is nothing but a spatial dispersion effect. Such effects are known to be tractable within electrodynamic theory, and indeed they have taken into account in an approach called ‘nonlocal electrostatics’6 (A. A. Kornyshev et al., 1978; M. A. Vorotyntsev, 1978; A. A. Kornyshev and M. A. Vorotyntsev, 1979) which generalizes 6 The

name ‘nonlocal electrostatics’ is unfortunate since it may create confusion by alluding to strange effects of ‘action at a distance’ and the like. The nonlocality we talk about is nothing of that sort, it just means that the dielectric response requires to take into account field values not at local points, but over a certain spatial range. All physical fields (potentials, electrostatic fields) remain well-defined local objects.



the commonly used electrodynamics of continuous media to take into account such spatial dispersion effects. The basis of this theory is the linear relationship between the dielectric displacement field D and the electric field E through a permittivity kernel which, in general, depends on two spatial arguments, Z 1 dr0 (r, r0 )E(r0 ) . D(r) = (4.55) 4π Here, (r, r0 ) is the dielectric permittivity tensor. It carries the new characteristic length scale, the correlation length λ of the water orientational polarization correlations introduced before. This length defines the scale for the deviation of the dielectric properties of a solvent from its average bulk value. For the protein geometry of Figure 4.3, the generalization of the Poisson equation in the solvent in nonlocal electrostatics reads as Z ∇ dr0 Σ (r, r0 )∇0 φΣ (r0 ) = 0 (4.56) Σ

where the primed symbol ∇0 denotes differentiation with respect to r0 . The main ingredient of the nonlocal theory is the integral kernel (r, r0 ) which contains the dependence on the correlation length λ. The mathematical expression for this model depends on the water model one wants to use. A simple, standard example for this quantity is the Lorentzian model7 |r−r0 |

Σ − ∞ e− λ (r − r ) = ∞ δ(r − r ) + . 4πλ2 |r − r0 | 0



As can be observed for this example - indeed this turns out to be of rather general nature - one can write eq.(4.57) in a more general form as (r − r0 ) = ∞ δ(r − r0 ) + ˜G(r − r0 )


where ˜ ≡ (Σ − ∞ )/4πλ2 , and G is a Green function satisfying LG = −δ(r − r0 )


with, in the given case, L ≡ ∆ − λ −2 , and G as the Green function of the Yukawa potential. Eq.(4.58) contains two terms; the first reduces to the local limit of the dielectric function ∞ at small distances r → r0 and is cancelled by the term ∝ ∞ in the second term. The remaining contribution is the local 7 The

Lorentzian model assumes isotropy of space as an additional simplifying assumption. Although this assumption does not hold strictly, it is an acceptable first approximation.


Computational Biology: A Statistical Mechanics Perspective

limit for large distances, Σ , the usual macroscopic dielectric constant. Eq.(4.58) can be taken as the starting point to reformulate the theory of nonlocal electrostatics as a local theory. For this, we in addition represent the dielectric displacement field as a sum of an irrotational part and a solenoidal part (a so-called Helmholtz decomposition), D(r) = −∇ψ(r) + ∇ × ξ(r) .


One can show that the solenoidal part, although it explicitly appears in the expression of the dielectric displacement field, does not appear in the equations of the electrostatic potentials (A. Hildebrandt, 2005). We are then finally left with the following system of equations for the electrostatic potentials φ and ψ, (A. Hildebrandt et al., 2004) ∆φΩ = −

4π % Ω


∆ψΣ = 0


Ω ∂n φΩ |Γ = ∂n ψΣ |Γ


φΩ |Γ = φΣ |Γ [∞ L − ˜] φΣ = 4πLψΣ (r) .

(4.64) (4.65)

Within its reformulation in terms of the local fields φ and ψ, the theory of nonlocal electrostatics can now be treated with standard approaches in order to numerically solve them with boundary element methods. Such a result is shown in Color Figure 5. In this figure, the local and the nonlocal electrostatic potentials of the enzyme trypsin are shown for comparison. It can clearly be seen that the structure of the electrostatic potential obtained in the nonlocal description deviates significantly from the local one which, for the same threshold value of the potential surfaces, hardly reaches beyond the geometric structure of the protein surface. From this application we can deduce that the electrostatic potential of proteins on ˚ Angstrom scales is markedly influenced by the water properties, and that electrostatic effects are important for the ‘visibility’ of a protein to its



interaction partners in solution.



Chromatin structure. In this section we break out of the central dogma of molecular biology, DNA → RNA → protein, which has so far been our guiding line. In any organism all these molecules have to act in concert, and the linear sequence should thus be better represented by a diagram with multiple feedback loops. In order to express genes, they first have to be switched on. This is done by specific proteins, the transcription factors in prokaryotes - and rather transcription factories, in eukaryotes. The function of these molecules requires recognition capability: proteins can detect their binding sites with high specificity. We will come back to this in Part III of the book when we look at the dynamics of transcription factor binding on DNA. In eukaryotes an additional problem arises, however. DNA is condensed and packed in the cellular nucleus, making a simple DNA-protein recognition process based on a diffusional search impossible. There must be a way for a protein or enzyme to find out where it has to go in the first place, even when the DNA is still condensed. Or, DNA has to be unpacked only partially, such that the search time for transcription factors becomes reasonably short. In order to understand the regulation of transcription in eukaryotes, we first have to get an idea of the compact form of DNA in the cell. The condensed form of DNA, in which it is actually a highly organized DNA-protein complex, is called chromatin. Chromatin is the structural basis of chromosomes. The different levels of organisation which are at present only partially understood. Our current knowledge is summarized in Color Figure 6 (from G. Felsenfeld and M. Groudine, 2003). The best understood element of chromatin structure is the nucleosome. A nucleosome is a DNA-protein complex which consists of an octameric protein complex built out of eight histone proteins (of four different kinds), around which DNA is wrapped along 146 bp with 1.75 left-winded turns. A linker histone at the exit of the DNA histones completes the structure; the precise positioning of the linker histone is not yet known. The molecular structure of the so-called nucleosomal core particle - the DNA wrapped around the histone octamer - has meanwhile been spatially resolved down to 1.9 ˚ A. A rib-


Computational Biology: A Statistical Mechanics Perspective

bon representation of a nucleosomal particle is shown in Color Figure 7 (from G. Felsenfeld and M. Groudine, 2003, after K. Luger at al., 1997). Unfortunately, the crystal structure of a nucleosome does not tell its full story. The N-terminal tails of the histone proteins are largely unstructured random coils, hence they do not crystallize. These tails are functionally important in two ways: • The tails are positively charged and can interact with the negatively charged DNA to form the condensed structures of chromatin, leading to both local or global structural changes of chromatin structure; • The tails are subject to various chemical modifications, brought about by specific enzymes. Some of these modifications can alter the charge of the chain. At the same time, many of the modifications can be read by specific proteins which affect chromatin structure locally. It has been argued that the histone modifications constitute a higher level regulatory code (B. D. Strahl and C. D. Allis, 2000). These two mechanisms are clearly not fully independent from each other, and are generally referred to as particular examples of chromatin remodelling. The notion of remodelling stands for the totality of dynamic structural changes chromatin can undergo and which are relevant for the regulation of transcription. Histone tail modifications. Enzymatic histone tail modifications are amino acid specific: not only the chemical nature of the amino acid is a determinant, but also its position on the tail. This is indicated in Figure 4.3. Several different types of modifications are presently known. The most important are: • Acetylation. One or several acetyl groups (CH3 CO) are transferred to lysine residues; the addition of one lysine reduces the positive charge of the tail by one. The modification is reversible, i.e., acetylating and deacetylating enzymes have been found. • Methylation. One (or more) methyl groups (CH3 ) are added to lysine residues; there is no effect on the charge. The reversibility of this modification is still under scrutiny; the first demethylating enzymes have recently been found, as discussed in A. Bannister and T. Kouzarides, 2005. • Phosphorylation. This modification can affect several residues: serine, threonine, tyrosine on the one hand, and lysine, histidine, and arginine on the other. The modifications are chemically different for both groups of amino acids.



FIGURE II.4.3: Specific enzymatic modifications on histone tails. (rib = ribosylation, is a rare modification.)

In general, histone tail modifications do differ somewhat between organisms, despite the high evolutionary conservation of the histones, their tails included. While the experimental evidence is growing that the histone tail modifications are read by a specific transcription machinery (T. Agalioti, G. Chen and D. Thanos, 2002), and apparently different modifications do not act independently (W. Fischle, Y. Wang and C. D. Allis, 2003), it is clear that the underlying ‘background’ mechanism is based on an electrostatic attraction of the negatively charged DNA and the positively charged histones. It is thus of interest to devise a simplified model view of this compaction mechanism. Chromatin fiber electrostatics. We here follow H. Schiessel, 2002. He considered the electrostatic interaction between two DNA strands at the entry-exit point of the DNA wrapped around the nucleosome, see Figure 4.4. The angle α between the two DNA strands defines the entry-exit angle at the nucleosome. For the geometry of Figure 4.4, it is defined by h0 (∞) ≡ tan(α/2) .


The relevant question to be answered is that of the dependence of the angle α on salt concentration cs , or on other charge-affected quantities. Salt contributions are relevant for the screening of electrostatic interactions, since the Debye screening length behaves as κ−1 = `D = (8πcs `B ) −1/2


where `B is the Bjerrum length, the measure for the respective importance of


Computational Biology: A Statistical Mechanics Perspective

FIGURE II.4.4: Schematic drawing of the entry-exit region around a nucleosome, after H. Schiessel, 2002. The two bent lines represent the lower and upper DNA strands; the entry-exit angle α is defined via the indicated asymptotics.

electrostatic forces and thermal energy we had introduced before in eq.(4.29). The DNA strands are considered as two semi-flexible polymers at positions ±h(x) with persistence length `P and a line-charge density λ. Their free energy is approximately given by  Z ∞  `P 00 2 2 F ≈ 2kb T dx (h ) + `B λ K0 (2κh(x)) (4.68) 2 0 where K0 (x) is the modified Bessel function of zero order.8 The first term in this expression is the bending energy of the strands, while the second describes their electrostatic interaction. The form of this term applies if it can be assumed that the interaction can be considered as that of a straight chain (hence the function K0 ) and a single charge at distance 2h. This approximation works as long as α is not too large; it tends to underestimate the true value.

8K α

is defined by Kα (x) =


π α+1 (1) i Hα (ix) 2


Hα (x) = Jα (x) + iYα (x) . Jα is the Bessel function we have introduced before; the Neumann function Yα (x) is defined by Jα (x) cos απ − J−α (x) Yα (x) = . sin αx



The Euler-Lagrange equation for h(x) reads, using the property K00 (x) = −K1 (x) of the Bessel function, `P h0000 = 2`B λ2 κK1 (2κh)


together with the boundary conditions h(0) = h0 (0) = 0 = h00 (∞) = h000 (∞) .


˜ = 2κh, x A variable change h ˜ = ((4lB λ2 κ2 )/lP ) 1/4 x allows to express eq. (4.69) as h0000 = K1 (h) (dropping tildes). With the relation s = h0 (x)|x=∞ one immediately obtains the dependence of the opening angle α on the physical parameters, in particular the length scales we defined in Section 4.3, s tan(α/2) = √ 2

`B `P

1/4  1/2 λ . κ


The value of the constant s in this expression is of O(1) (H. Schiessel, 2002). Solving for α, one finds the dependence α = arctan(C cs−1/4 ) .


Intramolecular electrostatic effects can also be included in the calculation. In this case one has to consider the modification of the persistence length `P by the charges along the chain: due to their mutual repulsion the charges along the molecule increase the rigidity of the chain. This additional repulsion can be calculated and leads to a modification of the persistence length `P , which has to be replaced by the Odijk-Skolnick-Fixman persistence length, ` = `P + `OSF = `P + `B λ2 /(4κ2 ) .


Plugging in numbers into the result eq.(4.72), one finds α-values of about 51o for 15 mM salt, and a value of 64o for 5 mM, in line with the expectation that a decrease of salt concentration favours the opening of the structure. The value of the entry-exit angle can be influenced by the histone tails if they bind to the DNA. This is a local effect, when the tails bind to the DNA attached to its own histone octamer. There is, however, also the important possibility that the tails of different nucleosomes interact. This gives rise to a tail-bridging effect (H. Schiessel, 2006), which is illustrated in Figure 4.5.


Computational Biology: A Statistical Mechanics Perspective

FIGURE II.4.5: A tail bridging two spherical particles and the interaction potential between them. After H. Schiessel, 2006.

Assuming that the nucleosomes can, in a zeroth-order approximation, as spherical particles of radius R, the interaction potential between them is given by a contribution of the repulsion between the two like-charges particles and the attraction caused by a bridge formed by a tail of contour length ` with the line density λ , U = Urep + Ubridge


which is explicitly given by (ignoring precise numerical prefactors)

`B Z 2 exp −d/`D U (d) = (4.75) kB T (1 + R/`D )2 d + 2R     `B `D Zλ `B `D Zλ − exp −(d − `D ) + exp −(` − ` ) . D R2 R2 Based on this result one can study how the minimum of the potential changes as the line charge is varied. The minimum distance of the two particles, dmin ∼ ln(C · λ), where C is a constant given by the other parameters of the potential. The potential minimum becomes very shallow for decreasing λ, and finally vanishes. Although these considerations give an idea about how electrostatic effects of the histone tails influence chromatin structure, there is still a large disparity between the understanding of any specific histone tail modification and the histone code hypothesis on the one side, and the physical compaction



mechanism on the other. Given the highly unspecific nature of the electrostatic interactions, and the apparent high specificity with which particular tail modifications can be placed and read, a satisfactory understanding of transcriptional regulation of chromatin seems still far out. In particular, it is yet open whether the histone tail specifications, highly specific as they may be, really constitute a code. Recent experimental work by M. F. Dion et al., 2005, on the acetylation of histone H4 provides evidence that only a very restricted number of these histone tail modifications are indeed ‘coding’ for specific regulatory processes. The conclusions support eearlier work by Q. Ren and M. A. Gorovsky, 2001, whose analysis of the acetylation patterns of a histone variant H2A.Z demonstrated the dominant role of rather non-specific electrostatic effects.


Bacterial chemotaxis: cooperativity once more

This final section in this Chapter on proteins considers a system which interacts cooperatively: a piece of the signal transduction pathway in bacteria. As it turns out, the modelling of the properties of this system needs both equilibrium and nonequilibrium methods. Although we limit ourselves here to the first approach, in some sense this section also prepares already for the models of dynamically interacting biomolecules of Part III of this book. But here we return, for the last time, to an example of the application of the Ising model in biology. How do bacteria sense molecules, e.g., to find a food source? They make use of protein assemblies in their membranes, built up from transmembrane receptors which recognise molecules in the environment. The receptor assemblies consist of well-known proteins and enzymes, a coupling protein CheW, and a histidine kinase, the enzyme CheA. In E. coli, these molecules are organized at the cell poles in tight clusters in thousands of copies. How do these clusters organize themselves to sense molecules of their environment? As it turns out, it is the coupling of neighbouring receptors which enhances the sensitivity with which signals can be detected. A physical model for the collective signalling of a receptor cluster has been built from the Ising model (Y. Shi and T. Duke, 1999). We assume that the state of a receptor molecule i can be characterized by a variable Vi , which can experimentally be related to a conformational state of the molecule. The signalling process occurs by the binding (docking) of a ligand which we describe


Computational Biology: A Statistical Mechanics Perspective

by a variable Bi . We can then write Vi = Vi ({Vj6=i } , {Bj }) .


where i, j = 1, 2, ..., N . Assuming for simplicity - supported by experimental evidence9 - that Vi has only two values Vi0 and Vi1 , we write for the state variable X Vi = ψ( Tij Vj + Bi − Ui )



where  ψ(x) =

V 1 , x > 0, V 0 , x ≤ 0.


In eq.(4.77), Tij is the nearest-neighbour coupling between two receptors, and Ui some threshold value. This type of states is known in the literature on neural networks as a McCulloch-Pitts threshold model. We identify V 0 with the active conformation; a value Bi > 0 then favors the binding of an attractant. With this construction, the receptor system is governed by the energy provided Tij = Tji and Tii = 0, H=−


Tij Vi Vj −


B i Vi +




Ui Vi .



This energy can be rewritten in Ising-spin notation (Exercise!) to yield H=−

X hiji

Jij Si Sj −


H i Si



where additional constant terms have been ignored. So we have arrived at the Ising model - but hold on - isn’t that a bit too crude a model? Indeed, while the coupling between the receptors, mediated by Jij seems very plausible, the ligand binding term merits some further interpretation. In fact, each receptor has two options: either a ligand is bound to it or it is not, and we decide that in the bound case, Hi = H, and otherwise Hi = 0. 9 The

two values of the state variable V correspond to different conformations of the protein, typically differently positioned helices.



That means we have to write down a bimodal probability distribution of the Hi : p(Hi ) = cδ(Hi − H) + (1 − c)δ(Hi ) ,


where c is a measure of the concentration of the liganded receptors. The average over the Hi is then given by Z H i = dHi Hi p(Hi ) = cH , (4.82) and the fluctuations of the random distribution of the Hi are p ∆Hi = c(1 − c)H .


We now solve this model in a mean-field theory approach and obtain m = tanh(βνJm + βHi ) ,


where the overline indicates that the tanh still needs to be averaged over Hi . Explicitly, one finds that m obeys a transcendental equation, m=

2c 2(1 − c) + − 1. 1 + exp[−2(βνJm + βH)] 1 + exp(−2βνJm)


Herein, ν is the number of nearest neighbors of the receptors, and β = 1/kB T . It is clear what these profiles will look like, since we are discussing an Ising model, after all: there will be two types of solutions generalizing the ferromagnetic and paramagnetic phases of the original Ising model. Of greater interest in the context of chemotaxis is the sensitivity to ligand docking, which we measure by the response of the order parameter m to a small concentration of ligands, hence S≡

(1 + exp(−2βH)) −1 − 1/2 1 ∂c m |c=0 = , 2 1 − βνJ


where m(c = 0) = 0 has been assumed. Obviously, S can be made arbitrarily large for βνJ → 1, which is exactly the value at the phase transition. We see that it is indeed the large cooperative effect which makes the system so sensitive. This model can be further extended in various ways, to include, e.g., a response to repellant molecules, to cover adaptation of the system to changes in environmental conditions, etc. We do not continue this here but refer to the literature, the paper by Shi and Duke, and others mentioned in the Additional


Computational Biology: A Statistical Mechanics Perspective

Notes. Instead, we want to finish this Part with a remark on what message should be taken along to the next. In this Part we have seen a multifaceted picture involving aspects of cooperative interactions between biomolecules and their relationship to phase transitions; we have seen the role played by disorder, and we have met non-specific charge-dependent effects. These three determinants of the static interactions of biomolecules will now find counterparts in dynamic effects. We will see structure formation driven by nonlinear interactions, but also disordering and ordering effects due to fluctuations in the stochastic dynamics of molecules.



Additional Notes Protein folding and docking. A modern standard reference on the protein folding problem is the book by A. Fersht, 2003. The CI2-molecule is discussed in some detail. There is an unconquerable amount of literature on protein folding. A recent listing of novel ideas is by D. J. Brockwell et al, 2000. The favourite approach within the statistical mechanics community is the development of as-simple-as-possible models to cover the large variety of observed structural features. A very recent prototypical example of this approach is the paper by T. X. Hoang et al., 2004. Note that the abstract begins “We present a simple physical model...”. Aspects of electrostatics for proteins are discussed in detail by B. Honig and A. Nichols, 1995. Electrostatics for protein docking is reviewed in A. Heifetz et al., 2002. A general review of recent protein docking methods is by G. R. Smith and M. J. E. Sternberg, 2002. Chromatin. A basic reference on the physics of chromatin is the review by H. Schiessel, 2003. A basic introduction to the biological problems associated with chromatin is the book by B. M. Turner, 2001. We particularly mention three open problems of chromatin physics and biology: Chromatin condensation. It is by now fairly clear that the mechanism underlying the condensation of chromatin is electrostatic in origin. The details of the process, and even more so, chromatin structure beyond the 10 nm fiber remain elusive so far. There has been some progress on the 30 nm fiber, in particular with sophisticated simulation methods J. Langowski and H. Schiessel, 2004, J. Sun et al., 2005. Nucleosome positioning. A problem of ongoing research on which much progress has been made recently is the position of the nucleosomes along the 10 nm fiber. Two aspects intervene here: the correlation of the positioning with the DNA sequence in contact with the nucleosome (M. S. Cosgrove et ´ et al., 2004), and the dynamic repositioning/remodelling al., 2004, M. Dlakic during transcriptional processes (H. Schiessel et al., 2001, H. Schiessel, 2006. Histone code and dynamics of regulation. A problem of outstanding interest in modern biology is how the histone and histone tail modifications are written and read in conjunction with the transcriptional states on the DNA. So far, there is no quantitative model, but a number of competing ideas


Computational Biology: A Statistical Mechanics Perspective

(S. Schreiber and B. E. Bernstein, 2002; A. Benecke, 2003, M. S. Cosgrove et al., 2004; W. Fischle et al., 2004, and G. Li and J. Widom, 2004.) Reviews of histone modifications are by B. M. Turner, 2000, S. L. Berger, 2002, and M. Iizuka and M. M. Smith, 2003. Bacterial chemotaxis. Bacterial chemotaxis is a problem which has been thoroughly investigated, both from a molecular as well as a ‘systems’ point of view. A lot of detail is known about the functional interactions between the receptors, see, e.g., the discussion by J. J. Falke, 2002, and the recent work by V. Sourjik and H. C. Berg, 2004. The latter paper presents a detailed discussion of molecular cooperativity in this system. A detailed kinetic model is by P. A. Spiro, J. S. Parkinson and H. G. Orthmer, 1997. The system is also a prime example of the notion of ‘robustness’ in the context of biological systems, which refers to the systems’ flexibility to respond to influences without requiring fine-tuning of system parameters. Seminal studies are N. Barkai and S. Leibler, 1997, and U. Alon et al., 1999.



References T. Agalioti, G. Chen, D. Thanos, Deciphering the transcriptional Histone Acetylation Code for a human gene, Cell 111, 381-392 (2002) U. Alon, M. G. Surette, N. Barkai and S. Leibler, Robustness in bacterial chemotaxis, Nature 397, 168-171 (1999) N. A. Baker, D. Sept, S. Joseph, M. J. Holst and J. A. McCammon, Electrostatics of nanosystems: Application to microtubules and the ribosome, Proc. Natl. Acad. Sci. USA 98, 10037-10041 (2001) A. Bannister and T. Kouzarides, Reversing histone methylation, Nature 436, 1103-1106 (2005) N. Barkai and S. Leibler, Robustness in simple biochemical networks, Nature 387, 913-917 (1997) A. Benecke, Genomic plasticity and information processing by transcription coregulators, ComPlexUs 1, 65-76 (2003) S. L. Berger, Histone modifications in transcriptional regulation, Curr. Op. Gen. Dev. 12, 142-148 (2002) D. J. Brockwell, D. A. Smith and S. E. Radford, Protein folding mechanisms: new methods and emerging ideas, Curr. Op. Struct. Biol. 10, 16-25 (2000) M. S. Cosgrove, J. D. Boeke and C. Wolberger, Regulated nucleosome mobility and the histone code, Nature Struct. Biol. 11, 1037-1043 (2004) B. Derrida, Random-Energy Model: Limit of a Family of Disordered Models, Phys. Rev. Lett. 45, 79-82 (1980) K. A. Dill and H. S. Chan, From Levinthal to pathways to funnels, Nat. Struct. Biol. 4, 10-19 (1997) M. F. Dion, S. J. Altschuler, L. F. Wu and O. J. Rando, Genomic characterization reveals a simple histone H4 acetylation code, Proc. Natl. Acad. Sci. USA 102, 5501-5506 (2005) M. Dlaki´c, D. W. Ussery and S. Brunak, DNA Bendability and Nucleosome Positioning in Transcriptional Regulation, in DNA Conformation and Transcription, Ch.14, T. Ohyama (ed.), (2004)


Computational Biology: A Statistical Mechanics Perspective

J. J. Falke, Cooperativity between bacterial chemotaxis receptors, Proc. Natl. Acad. Sci. USA 99, 6530-6532 (2002) G. Felsenfeld and M. Groudine, Controlling the double helix, Nature 421, 448-453 (2003) A. Fersht, Structure and Mechanics in Protein Science, W. H. Freeman (2003) W. Fischle, Y. Wang and C. D. Allis, Binary switches and modification cassettes in histone biology and beyond, Nature 425, 475-479 (2003) A. Heifetz, E. Katchalski-Katzir and M. Eisenstein, Electrostatics in proteinprotein docking, Protein Science 11, 571-587 (2002) A. Hildebrandt, R. Blossey, S. Rjasanow, O. Kohlbacher, H.-P. Lenhof, Novel formulation of nonlocal eletrostatics, Phys. Rev. Lett. 93, 108104 (2004) A. Hildebrandt, Biomolecules in a structured solvent, Rhombos (2005) T. X. Hoang, A. Trovato, F. Seno, J. R. Banavar and A. Maritan, Geometry and symmetry presculpt the free-energy landscape of proteins, Proc. Natl. Acad. Sci. USA 101, 7960-7964 (2004) B. Honig and A. Nichols, Classical Electrostatics in Biology and Chemistry, Science 268, 1144-1149 (1995) M. Iizuka and M. M. Smith, Functional consequences of chromatin modifications, Curr. Op. Gen. Dev. 13, 154-160 (2003) J. D. Jackson, Classical Electrodynamics 3rd ed., Wiley (1999) R. M. Jackson and M. J. E. Sternberg, A Continuum Model for ProteinProtein Interactions: Application to the Docking Problem, J. Mol. Biol. 250, 258-275 (1995) J. Janin and S.J. Wodak, Computer analysis of protein-protein interactions, J. Mol. Biol. 124, 323-342 (1978) E. Katchalski-Katzir, I. Shariv, M. Eisenstein, A. A. Friesem, C. Afaflo, I. A. Vakser, Molecular Surface Recognition: Determination of Geometric Fit Between Proteins and Their Ligands by Correlation Techniques, Proc. Natl. Acad. Sci. USA 89, 2195-2199 (1992)



A. A. Kornyshev, A. I. Rubinstein and M. A. Vorotyntsev, Model nonlocal electrostatics: I, J. Phys. C: Solid State Phys. 11, 3307-3322 (1978) A. A. Kornyshev and M. A. Vorotyntsev, Model nonlocal electrostatics: III. Cylindrical interface, J. Phys. C: Solid State Phys. 12, 4939-4946 (1979) J. Langowski and H. Schiessel, Theory and computational modelling of the 30 nm chromatin fiber, New Comprehensive Biochemistry 39, 397-420 (2004) C. Levinthal, How to fold graciously, in M¨ ossbauer Spectroscopy in Biological Systems, P. Debrunner and J. C. M.Tsibris (eds.), 22-24 (1968) G. Li and J. Widom, Nucleosomes facilitate their own invasion, Nat. Struct. Mol. Biol. 11, 763-769 (2004) K. Luger, A. W. M¨ ader, R. K. Richmond, D. F. Sargent and T. J. Richmond, Nature 389, 251-260 (1997) G. S. Manning, Limiting Laws and Counterion Condensation in Polyelectrolyte Solutions I., J. Chem. Phys. 51, 924-933 (1969) C. Merlo, K. A. Dill and T. R. Weikl, Φ-values in protein-folding kinetics have energetic and structural components, Proc. Natl. Acad. Sci. USA 102, 10171-10175 (2005) J. N. Onuchic, Z. Luthey-Schulten and P. G. Wolynes, Theory of Protein Folding: The Energy Landscape Perspective, Annu. Rev. Phys. Chem. 48, 545-600 (1997) Q. Ren and M. A. Gorovsky, Histone H2A.Z Acetylation Modulates an Essential Charge Patch, Molecular Cell 7, 1329-1335 (2001) H. Schiessel, J. Widom, R. F. Bruinsma and W. M. Gelbart, Polymer Reptation and Nucleosome Repositioning, Phys. Rev. Lett. 86, 4414-4417 (2001) H. Schiessel, How short-ranged electrostatics controls the chromatin structure on much larger scales, Europhys. Lett. 58, 140-146 (2002) H. Schiessel, The physics of chromatin, J. Phys. Cond. Mat. 15, R699R774 (2003) H. Schiessel, The nucleosome: A transparent, slippery, sticky and yet stable DNA-protein complex, Eur. Phys. J. E 19, 251-262 (2006)


Computational Biology: A Statistical Mechanics Perspective

S. Schreiber and B. E. Bradstein, Signaling Network Model of Chromatin, Cell 111, 771-778 (2002) Y. Shi and T. Duke, Cooperative model of bacterial sensing, Phys. Rev. E 58, 6399-6406 (1998) T. Simonson, Macromolecular electrostatics: continuum models and their growing pains, Curr. Op. Struct. Biol. 11, 243-252 (2001) G. R. Smith and M. J. E. Sternberg, Prediction of protein-protein interactions by docking methods, Curr. Op. Struct. Biol. 12, 28-35 (2002) V. Sourjik and H. C. Berg, Functional interactions between receptors in bacterial chemotaxis, Nature 428, 437-441 (2004) P. A. Spiro, J. S. Parkinson and H. G. Othmer, A model of excitation and adaptation in bacterial chemotaxis, Proc. Natl. Acad. Sci. USA 94, 72637268 (1997) B. D. Strahl and C. D. Allis, The language of covalent histone modifications, Nature 403, 41-45 (2000) J. Sun, Q. Zhang and T. Schlick, Electrostatic mechanism of nucleosomal array folding revealed by computer simulation, Proc. Natl. Acad. Sci. USA 102, 8180-8185 (2005) B. M. Turner, Histone Acetylation. A global regulator of chromatin function, in C. Nicolini (ed.), Genome Structure and Function, Kluwer Academic Publishers, 155-171 (1997) B. M. Turner, Chromatin and gene regulation, Blackwell Science (2001) B. M. Turner, Cellular Memory and the Histone Code, Cell 111, 285-291 (2002) M. A. Vorotyntsev, Model nonlocal electrostatics: II, Spherical interface, J. Phys. C: Solid State Phys. 11, 3323-3331 (1978) P. Wolynes, Landscapes, Funnels, Glasses and Folding: From Metaphor to Software, Proc. Americ. Philosoph. Soc. 145, 555-563 (2001)

Part III


Chapter 1 Network Dynamics I: Deterministic

How come all my body parts so nicely fit together? All my organs doing their jobs, no help from me!

Crash Test Dummies, How does a duck know? (1993)

The detailed knowledge of the properties of biomolecules and their interactions, as we have described them in Part II of this book, is a necessary but not a sufficient step to understand biological systems. The next step is to understand how the different molecular components interact in concert and how they build up the hierarchy observed in biological systems: cellular compartments, cells, organs and organisms. In most of what follows, we will be interested here in genetic networks of some sort. Under this term we understand biological reaction schemes on the basis of the central dogma, DNA makes RNA makes protein. We will encounter the following situations: • A protein attaches itself to DNA at a binding site; such molecules are transcription factors which help control gene expression; • A protein-RNA complex, the RNA polymerase, reads out the genes after fixing itself at a promoter site; the readout is a messenger RNA; • the mRNA transcript is read at a ribosome and serves as a blueprint for a protein; • proteins compete with each other; this may happen by their direct interaction or blocking or activating the corresponding genes. This basic readout mechanism underlies the complexity of all gene networks. The examples we discuss have been selected to allow for - mostly - analytical calculations to illustrate several aspects arising in systems biology.



Computational Biology: A Statistical Mechanics Perspective

We will build this up in a step-wise fashion and progress through a description of a simple regulatory circuit, modelled by deterministic differential equations and move on to the emergence of spatial structures. This material is covered in Chapter 1. Chapter 2 is concerned with the effects of fluctuations on the regulatory components. Finally, in Chapter 3, we will take a more global view and have a look at the overall structure of biological networks.1


Deterministic dynamics: λ-repressor expression

In this section we want to discuss an example of a biological network - a very small one - and to formulate a simple model for a particular aspect of its dynamics. The example is the λ-phage. The λ-phage is a bacterial virus for which the molecular basis of its life cycle is very well understood (M. Ptashne, 2004). Consequently, it is often used as a prototypical example for theoretical studies. Here, we use it as a ‘simple’ complex system to illustrate how to describe biological networks within a deterministic dynamics setting. But first, we need to learn some basics of the life cycle of the virus; many more details can be found in Ptashne’s book. Life cycle of phage λ. Phage λ is a bacterial virus which infects E. coli. The virus inserts its DNA into the bacterial cell. Here, the DNA has two options. The first is its insertion into the host genome, and it then ‘stays on board’ as a ‘blind passenger’, being passed on from one generation to the next upon cell division. This phase is called the lysogenic phase. But if, e.g., the bacterium is endangered and responds to this external stress, the virus can ‘abandon ship’. Its genome is excised out of the host genome and begins the production of new viruses which then leave the host (which dies). The phase in which the virus replicates is called the lytic phase. The decision making process between the lysogenic or lytic phase is made at a molecular switch. The switch is illustrated schematically in Figure 1.1. It consists of an operon built from three operator binding sites, at which a dimer of the λ-repressor molecule can attach itself. The λ-repressor is a transcription factor, i.e., its presence on DNA serves as a flag for RNA polymerase to attach itself and start transcription of an adjacent gene. Indeed, the three 1 The

insider immediately recognizes this progression as typical for the statistical physics approach: for few components, detailed models can be built - while things again simplify when a large number of components can be studied. Admittedly, the real difficulty lies in the ‘middle thing’.

Network Dynamics I: Deterministic


FIGURE 1.1: The λ-phage operon. Operator sites are shown as grayshaded areas, with OR3-OR2-OR1 from left to right. Promoters are indicated by dashed frames; gene transcription directions are indicated by arrows. Two λ-repressor dimers are shown upon attachment and approach. (After M. Ptashne, 2004.)

operator sites are at the same time overlapping with two neighbouring promoter regions, which are the fixation platforms for RNA polymerase, and they directly neighbour two adjacent genes, cI, the gene for the λ-repressor molecule, and cro, the gene for a second transcription factor also involved in the switch. Although operator and promoter sites are overlapping - and part of their function depends on this fact - their biological role is different. The switch now functions as follows. The attachment of a CI-dimer (the other name of the dimer form of the λ-repressor) at operator site OR1 enhances the fixation probability of a second dimer to fix at OR2. The fixation of CI at OR2, in turn, increases the probability of fixation of RNA polymerase at the cI-promoter, upon which transcription of cI occurs. The action of cI is thus directly implied in its description, hence the gene is autoregulatory. The continual transcription of cI ensures that the system stays in the lysogenic phase. If, on the other hand, the system needs to switch to the lytic phase, a repressor dimer can fix at OR3, which due to its overlap with the cI-promoter blocks the access for RNA-polymerase. Now, Cro-proteins intervene, and they attach themselves to the same OR-sites as CI, but with inverted affinities. A Cro-protein present at OR2 enhances the fixation probability for a polymerase at the cro-promoter, and transcription of cro can start, initiating the lytic pathway. Let’s now build some of these mechanisms into a mathematical model.


Computational Biology: A Statistical Mechanics Perspective

Hasty model I: basics. In this subsection we discuss a simple model for repressor expression proposed by J. Hasty et al., 2000. The main actor in this model is the repressor molecule itself; we denote it by X. The advantage of the Hasty model is that it allows to quickly gain a qualitative insight, but it can be made fully quantitative as well. Here, we are mainly interested in the qualitative aspects, though. A first simplification in the Hasty model is the assumption that one can neglect the operator binding site OR1. This is justified if one is only interested in the repressor kinetics: it means we do not want to model the full switch, but just the autoregulation mechanism of the λ-repressor. In this case, the system can be described by six reactions. Four of them will be assumed as being fast when compared to the others; these are the association and dissociation of the transcription factors from their binding sites. They will be considered as equilibrium reactions, and are given by • Repressor dimerization, 2X ↔ X2 with equilibrium constant K1 ; • Repressor dimer binding to the DNA promoter site OR2 on D, D+X2 ↔ DX2 , with equilibrium constant K2 ; • Repressor dimer binding to the DNA promoter site OR3 on D, D+X2 ↔ DX2∗ , with equilibrium constant K3 ; • Repressor dimer binding to the DNA promoter sites OR2 and OR3 on D, DX2 + X2 ↔ DX2 X2 , with equilibrium constant K4 . The slow reactions in the system are transcription and degradation. Transcription is described by the irreversible reaction scheme DX2 + P →kt DX2 + P + nX


in which P is the DNA polymerase, and n is the number of repressor proteins per RNA transcript. The degradation reaction reads X →k2 φ .


We now define concentrations for all variables, x ≡ [X] , y ≡ [X2 ] , d ≡ [D] , u ≡ [DX2 ] , v ≡ [DX2∗ ] , z ≡ [DX2 X2 ] . (1.3) Considering the fast reactions as equilibrium reactions allows to write them

Network Dynamics I: Deterministic


as simple algebraic expressions, y u v z

= K1 x2 = K2 dy = K1 K2 dx2 = σ1 K2 dy = σ1 K1 K2 dx2 = σ2 K2 uy = σ2 (K1 K2 )2 dx4 .

(1.4) (1.5) (1.6) (1.7)

It now remains to write down a rate equation for the repressor, which is given by x˙ = α ˜ u − kd x + r . (1.8) Here, α ˜ = nkt p0 is a constant containing kt and the concentration of polymerase, p0 , which is assumed fixed. The term −kd x describes degradation of x while the last term, r, constitutes a basal transcription rate. In order to close the system of equations, a conservation law needs to be invoked. In fact, the total concentration of DNA promoter sites, dT , is fixed.2 Thus dT = d + u + v + z = d(1 + (1 + σ1 )K1 K2 x2 + σ2 K12 K22 x4 )


which leads to d=

dT . 1 + (1 + σ1 )K1 K2 x2 + σ2 K12 K22 x4


This use of expression eq.(1.10) allows us to rewrite eq.(1.8) in a succinct form containing only x as a variable, x˙ =

αx2 − γx + 1 , 1 + 2x2 + 5x4


where repressor concentration and time have been rescaled to dimensionless quantities, and the values σ1 ≈ 1, σ2 ≈ 5, obtained from experimental estimates, have been used. The remaining parameters α and γ then constitute ratios of transcription rate and degradation rate relative to the basal transcription rate. The mathematical discussion of this equation is extremely simplified by its one-dimensional character. In fact, we can rewrite it in the form x˙ = −∂x V (x) 2 This


is true in vivo, for small concentrations, but even more so in vitro, for large concentrations.


Computational Biology: A Statistical Mechanics Perspective

where V (x) is the integral of the right hand side of eq.(1.12). V (x) is now seen to serve as a ‘potential energy’ landscape, shown in Figure 1.2, a function which has two minima of different depth. Whichever is the lower of the two determines the steady state of the repressor system, i.e., either a state with a low concentration, or with a high concentration of repressor molecules.3 Figure 1.3 is the main result of this subsection; it shows that repressor expression is a bistable system arising from the two competing mechanisms of repressor production and degradation.

FIGURE 1.2: The effective potential of the λ-repressor. From J. Hasty et al., Copyright (2000) National Academy of Sciences, U.S.A.

Hasty model II: looping included. We now include looping into the model. What this means is indicated in Figure 1.4. In the λ-phage a second, or left, operator region (OL) is found along the genome. This region, to which the corresponding genes are lacking, has a peculiar regulatory function ´vet et al., 1999). Repressors that fix themselves at the OL-sites, are (B. Re present to interact with their cousins sitting at OR-sites by forming a repressor octamer. Octamerization of repressors in solution is a very rare event, but the presence of the molecules on the DNA increases the formation rate enormously, since looping is a very effective mechanism of bringing different regions of DNA into contact (H. Merlitz et al., 1998). It is therefore also a very common regulatory mechanism in eukaryotes. 3 This

argument, as nice as it is, works indeed in general only for one variable. For more than one variable, a potential will only exist in very special cases.

Network Dynamics I: Deterministic


FIGURE 1.3: Bistability in phage λ. Graph A shows the graphical solution of the stationary points of eq.(1.12), i.e., x˙ = 0. In the figure, the solid curve is the first term of the rhs of eq.(1.12), while the broken lines are given by the function f (x) = γx − 1. The intersections of the graphs correspond to the extrema of V (x), and can, e.g., be varied by changing γ. Graph B shows a hysteresis diagram of repressor concentration vs. γ. The size of the hysteresis loop is a simple criterion for the barrier between both states: the bigger the loop, the higher the barrier. The non-mutant case refers to the situation in which all operator sites are present, while the mutant case corresponds to the simplified model we discussed in the text. Qualitatively, both cases are similar, but the quantitiative difference is significant. From J. Hasty et al., Copyright (2000) National Academy of Sciences, U.S.A.


Computational Biology: A Statistical Mechanics Perspective


Looping in phage λ. After M. Ptashne, 2004.

The Hasty model can be extended to account for DNA looping; one needs two additional assumptions. Again we simplify the structure of the left operator as we did for the right: • the second operator unit equals the first: OR = OL; this in particular refers to all rate constants. OL is placed on a second DNA site, dL . • there is an additional kinetic process described by the complex formation between a doubly bound repressor at OR and the same at OL. This process has a rate K5 , and is assumed to be in equilibrium (fast compared to repressor production). The presence of this additional mechanism modifies the nonlinearity in eq. (1.12), which we now call f` (x), in the case looping. The change is due to the different possibilities of repressor fixation at either OR or at OL (dR or dL ). Assuming the symmetric case with dR = dL = d` /2 we find the conservation law δd2 dT = d` (1 + (1 + σ1 )x2 + σ2 x4 ) + ` x8 (1.13) 4 which replaces eq.(1.9), and where the factor δ derives from the looping process, i.e., the kinetic rate K5 . When the equation is solved for d` this yields, with σ1 = 2 and σ2 = 5, #   "  1/2 1 + 2x2 + 5x4 dT δx8 d` = 2 1+ −1 . (1.14) δx8 (1 + 2x2 + 5x4 )2 Although expression (1.14) looks rather different from (1.10), there is in fact only a minor difference between them. This fact becomes obvious by expanding the square root in eq.(1.14) for small x, which obviously reduces (1.14) to

Network Dynamics I: Deterministic


the eq.(1.10), while for large x the curves have the same asymptotics but a different amplitude. This insensitivity to the presence of the nonlinear coupling is astonishing, but it allows to understand the role played by the looping-mediated coupling for the repressor dynamics. If one compares the stationary points of the repressor system with and without looping one sees that the effect of looping is simply to reduce the bistable region in which repressed and non-repressed transcriptional states coexist, see Figure 1.3. Given that the full operator region indeed has a significantly larger bistable region due to the presence of the operator site OR1 which we neglected in the calculation, the looping mechanism can be understood as an effective means to dynamically modulate the bistability of the repressor autoregulation. As a further consequence, the looping mechanism also affects the fluctuations, which we have not discussed at all. Since, with looping allowed for, the system can switch between the two states within a much narrower concentration range, the frequency of switching will consequently increase, while the amplitude diminishes. This is in accord with recent conclusions based on stochastic simulations (J. M. G. Vilar and S. Leibler, 2005). It is worth noting that the simple model presented here neglects all spatial structure of the looping mechanism; the mechanism appears identical to that of an octameric complex binding at OR. But it is this process which is highly unlikely to occur since it has a very small equilibrium constant. By contrast, bringing dimer complexes at sufficiently distant sites into contact via looping occurs with a higher probability. The binding of repressor dimers at the OL region thus constitutes an effective mechanism to hold the molecules ‘in stock’ by placing them on the DNA, rather than letting them diffuse through solution.



Computational Biology: A Statistical Mechanics Perspective

The Turing insight

However, a partial differential equation will be obtained which is thought to give a good approximation to the behaviour of certain kinds of chemical system. The differential equation has a number of parameters and it is necessary to find values for these parameters which will make the differential equation behave properly. Alan Turing, The development of the daisy (1952)

In the 1950s Alan Turing hypothesized that the formation of biological structures may be driven by chemical instabilities. He proposed that the patterning results from an instability involving the reaction and diffusion of particular molecules, the morphogens. The idea that structure formation is based on the spatial distribution of chemical compounds and their interactions has proved to be a very profound one; meanwhile, several different variations of this scheme exist. We will go through the most important ones. Morphogen gradients. The simplest way to generate a pattern is to assume that a morphogen is present at some localized source, and then diffuses into the surrounding tissue. The simplest chemical reaction that can happen is the decay of the morphogen due to degradation. Such a system is described by the differential equation ∂t φ = D∂x2 φ − µφα + ηδ(x) ,


where φ is the concentration of the morphogen, D is its diffusion coefficient in the tissue, and µ the degradation coefficient. The delta-function term describes the localized source of the morphogen. Note that we have restricted ourselves to a one-dimensional case. The parameter α allows to tune the degradation process: for α = 1 one has the ‘simple’ degradation process while the value α = 2 corresponds to an enhanced autodegradation. This case has recently been studied by (J. L. England and J. Cardy, 2005). Eq.(1.15) then has a steady-state solution φ0 (x) given by 6D 1 φ0 (x) = (1.16) µ (x + )2 where  ≡ (12D2 /µη)1/3 . This solution corresponds to a gradient profile which decays algebraically outwards from the source.

Network Dynamics I: Deterministic


A morphogen gradient can serve as a very simple means to structure space. Any mechanism capable to detect absolute concentrations, say a value φ∗ , naturally devides space into regimes φ > φ∗ and φ < φ∗ . Although this mechanism seems astonishingly simple, it is actually realised in many systems (see, e.g., J. B. Gurdon and P. Y. Bourillot, 2001). Activators and inhibitors. More complex structuring effects can arise by considering not only one substance which decays, but also the interaction of two substances, one of which we call ‘activator’ a, while the second is called an ‘inhibitor’, h. A possible set of reaction-diffusion equations for such a system is given by %a2 ∂t a = Da ∂x2 a − µa + + %a , (1.17) h (1.18) ∂t h = Dh ∂x2 h − νh + %a2 + %h , where the terms ∝ %a2 describe the auto- and crosscatalysis of the activator as opposed to an autodegradation we had in eq.(1.15) - while the term ∝ 1/h covers the action of the inhibitor; the remaining terms are degradation and source terms, as one verifies upon comparison with eq.(1.15). Equations (1.17), (1.18) are examples of a large class of models which were shown to generate patterns by a combination of short-range activation due to the catalytic nonlinearity ∝ a2 and a long-range inhibition. Several examples with different types of nonlinearities are discussed in the seminal paper by (A. Gierer and H. Meinhardt, 1972). We will come back on the choice of nonlinearities later on. We now turn to the technical discussion of the linear instability in the reaction-diffusion systems. The Turing model. Turing was the first to see that instabilities in reaction-diffusion systems can lead to a pattern formation process. Let us discuss the appearance of an instability in the reaction-diffusion equations. Suppose there are two chemical substances of concentrations φ1 and φ2 , each of which is allowed to react according to a chemical reaction scheme (see Part I, Chapter 2). As before, the substances may diffuse. We now write down a general system for the reaction-diffusion equations ∂t φ1 = f (φ1 , φ2 ) + D1 ∆φ1 ∂t φ2 = g(φ1 , φ2 ) + D2 ∆φ2


wherein f and g are nonlinear functions of φ1 and φ2 , D1,2 are diffusion con-


Computational Biology: A Statistical Mechanics Perspective

stants, and ∆ ≡ (∂x2 +∂y2 +∂z2 ) is the Laplace operator in cartesian coordinates in three dimensions, generalizing what we wrote down before in eqs.(1.17), (1.18). Suppose now that this system of equations has a stationary point, ∂t φ1 = ∂t φ2 = 0 at (φ1 , φ2 ) = (0, 0). A linearization of eqs.(1.19) around this state in the form φ = φ0 + δφ then leads to a linear system in the fluctuations δφ; we drop the δ in the following to simplify notation: d dt

φ1 φ2


a11 a12 a21 a22

     φ1 D1 ∆φ1 · + . φ2 D2 ∆φ2


The coefficients aij will in general depend on the value of φ0 . Now suppose that the system is confined to a cubic volume. Assuming that there is no flux of φ1 or φ2 entering the cube of side length `, a Neumann boundary condition holds, i.e., the first spatial derivative of φ1,2 with respect to the wall normal vanishes. The solutions of eqs.(1.20) can then be written in a general form as     X 2π 2π x1 · · · cos x3 φ1 (x, t) = cn1 ,...,nk (t) cos ` ` n1 ,n2 ,n3 ≥0     X 2π 2π dn1 ,...,nk (t) cos φ2 (x, t) = x1 · · · cos x3 ` ` n1 ,n2 ,n3 ≥0

(1.21) where c.. , d.. are the Fourier coefficients of the solution to eqs.(1.20). The terms in the sum are the eigenmode solutions, indexed by the tuple set N ≡ (n1 , n2 , n3 ). The introduction of these expressions into eqs.(1.20) then leads to an infinite system of ordinary differential equations for the Fourier coefficients, given by d dt

cN dN


2 2 a12 a11 − ( 2π ` ) D1 |N | 2 2 a21 a22 − ( 2π ` ) D1 |N |

   cN · (1.22) dN

where |N |2 = n21 + n22 + n23 . We denote, in what follows, the (2 × 2)-matrix in eq.(1.22) by MN . For each non-negative 3-tuple N the stability of the eigenmodes is determined by the eigenvalues of MN , λ± N . These can be expressed in terms of the trace and the determinant of the matrix MN ,  p 1 2 − 4DetM λ± = T rM ± (T rM ) . (1.23) N N N N 2

Network Dynamics I: Deterministic


For every N ≥ (0, 0, 0) the real and imaginary parts of eq. (1.23) are bounded from above; the value Λ ≡ max Re λ± N


thus is the upper bound of the spectral abscissae of the set of matrices {MN } (in the positive quadrant). We see already by inspection of the entries of the matrix MN that the diffusion terms may indeed destabilize a stable state of the homogeneous system, which may lead to a symmetry breaking of the global behaviour of the solutions to the nonlinear equations. As it turns out, for the generation of patterns, the two diffusion constants must significantly differ from each other, which may not be evident in many biological situations: why should two proteins that, say, diffuse in the cytoplasm have significantly different diffusion constants? One way out, as we will see below, is that one diffusion process is confined to the cell membrane, while the other occurs in the cytoplasm. Returning to the eigenvalue problem of the Turing system, there are two types of eigenvalues that indicate the linear instability of the homogeneous state: in case of a Turing instability, Λ is positive and is one of the eigenvalues of MN ; if there is an oscillatory instability, there exists a mode with ± Re λ± N = Λ > 0, but for which Im λN 6= 0. The occurrence of a Turing instability is, however, not generally sufficient to generate a Turing pattern, i.e., a time-independent spatial structure. We do not follow this trace here, since it is not necessarily a stationary pattern that is of interest to us in the description of spatial structures in biological systems. Also, patterning may arise in systems in which the conditions built into, e.g., the Gierer-Meinhardt model, like the production of an activator, need not be the case. We will now see a particular example, the Min system, to which this applies.


The Min system

In this section, we study an example of a pattern forming processes: the cell division mechanism in E. coli. E. coli is a rod shaped bacterium of 2-6 µm length and 1-1.5 µm diameter. It divides roughly every hour: after replication of the DNA, a cell splits in half to form two daughter cells. The question is: how does the cell ‘know’ where its middle is?


Computational Biology: A Statistical Mechanics Perspective

The mechanism for cell division is based on the action of three proteins, the Min family: MinC, MinD and MinE. These proteins oscillate from one end to the other within the cytoplasm, and, furthermore, they move between the cytoplasm and the cytoplasmic membrane. Their concentrations determine the site for the cell division to occur. The cell division itself is brought about by the formation of a contractile polymeric ring, called Z-ring, at the site of division. This ring is made from a protein named FtsZ, which is a homologue of tubulin which plays a similar role in higher organisms. FtsZ forms just underneath the cytoplasmic membrane, but how it generates the force needed for cell devision is still unknown. The Min-system serves to confine the location of this ring to midcell. If this mechanism is perturbed, ring formation can occur in the vicinity of the cell poles, and can lead to the formation of minicells devoid of DNA. The action of the Min-system, as far as it is now known from molecular biology studies based on deletion mutants (i.e., bacteria in which corresponding genes have been knocked out), is as follows: • MinC localized in the cytoplasmic membrane locally inhibits assembly of the contractile Z-ring. It remains in the cytoplasm and inactive in the absence of MinD; • MinD binds MinC and recruits it to the cytoplasmic membrane; • MinE drives MinD away from the membrane; the fact that it is driven mostly away from the bacterial midplane and hence allows for ring formation is a result of the protein dynamics. Without MinE, the MinC/MinD couple would inhibit Z-ring formation everywhere, blocking cell division. In this case, long filamentous cells are observed. Without MinC, Z-ring formation cannot be inhibited anywhere, leading to inviable cells. Without MinD, neither MinC nor MinE are recruited to the cytoplasmic membrane and hence have less effect. The diffusion dynamics of the Min-system has been visualized by fluorescence techniques. Hereby, the Min proteins are modified to fusion proteins with GFP (green fluorescent protein). This is illustrated in Figure 1.5. The MinC/Min D accumulate at a polar end of the bacterium on the cytoplasmic membrane. MinE forms a band at midcell which sweeps to the polar end and ejects the MinC/MinD into the cytoplasm. The ejected proteins then rebind at the other end of the bacterium. The MinE band, by contrast, dissociates at the pole and reforms at the center, and the whole process repeats towards the opposite cell pole. The observed oscillation period lasts 1-2 min, well below the cell division cycle length. The net effect of the dynamics is to minimize the concentration of MinC/MinD at midcell, and maximizing MinE concentration there, see Figure 1.6. The result shown there is obtained from a model

Network Dynamics I: Deterministic


FIGURE 1.5: MinD oscillations in E . col i . The pictures a)-d) show fluorescence images of MinD-GFP at subsequent time points separated by 20 s; the scale bar in a) is 1 µm. Picture e) shows a time-average of frames separated by 1s during one full oscillation cycle. Finally, Picture f) shows the fluorescence intensity I from a line scan of the time under e). The total cell length is Lc = 2.3 µm. (Reprinted with permission from G. Meacchi and K. Kruse, 2005.)

developed by M. Howard et al., 2001, to which we will now turn.

The quantitative model for this dynamics based on the reaction-diffusion equation proposed by M. Howard et al., 2001, is based on four coupled equations for the densities of MinD and MinE in the cytoplasm and the cytoplasmic membrane. Since the MinC dynamics follows that of MinD, it does not need to be considered explicitly. The equations read (here, densities are %i with i = e, E and i = d, D for the membrane (small) and cytoplasm (capital)):


Computational Biology: A Statistical Mechanics Perspective

FIGURE 1.6: Concentration profiles of the Min proteins in E . col i . (M. Howard et al.; Copyright (2001) American Physical Society)

σ1 %D + σ 2 %e %d 1 + σ10 %e σ 1 %D ∂t %d = − σ 2 %e %d 1 + σ10 %e σ 4 %E ∂t %E = DE ∂x2 %E − σ3 %E %D + 1 + σ40 %D σ 4 %e ∂t %e = σ 3 %E %D − 1 + σ40 %D

∂t %D = DD ∂x2 %D −

(1.25) (1.26) (1.27) (1.28)

The first reaction terms in the equations for MinD describe the spontaneous association of MinD to the cytoplasmic membrane. Cytoplasmic MinD recruits cytoplasmic MinE to the membrane via the reaction term in the equation for MinE. Once there, MinE drives MinD into the cytoplasm, given by the second reaction term in the equations for MinD. The last term in the equations for MinE corresponds to the spontaneous dissociation from the membrane. All these terms have been modelled by introducing a parameter set {σ}, with values given below. This dynamics conserves protein number; there are no source or drain terms. The model exhibits a linear Turing-like Hopf-instability. Note that the mechanism of pattern formation in this system differs from the one proposed by Turing, and Meinhardt and Gierer, since in this system there is no production of activators or an external influx of particles. The origin of the instability is the disparity between the diffusion rates between the cytoplasm and the cytoplasmic membrane; this has been idealized by setting the diffusion rates in the membrane to zero. Task. Analyze the linear stability of the system, testing for solutions of the

Network Dynamics I: Deterministic

TABLE 1.1:


Min system parameters

DD (µm2 /s) DE (µm2 /s) σ1 (s−1 ) σ10 (µm) 0.28




σ2 (µm/s)

σ3 (µm/s)

σ4 (s−1 ) σ40 (µm)





form eλt+iqx . Take the parameter values used by M. Howard et al., 2001, listed in Table 1.1. The equations (1.25) are valid only in a qualitative way, but establish a general basis for further improvement, which has occurred over the last years. But it should be clear that there is no general criterion to precisely decide on the form of the nonlinearities - instead, comparisons with different experiments need to be made in order to improve many details of the model. Task. Discuss the difference between the models proposed by M. Howard et al., and H. Meinhardt and P. A. J. de Boer, both 2001.


Computational Biology: A Statistical Mechanics Perspective

Additional Notes Deterministic modeling. The modelling of the dynamics of networks by nonlinear differential equations has a long history. Two of the pioneers are A. Goldbeter and J. J. Tyson, see their work listed in the references. The lambda phage. The lambda phage is as well a superb model system for biologists, as a system to model for the biophysicist. The first modelling attempt, based on the free energies of binding of the transcription factors and, in particular, their cooperative behaviour, is by G. Ackers et al., 1982. Stochastic analyses were performed by A. Arkin et al., 1998 - but this is actually a topic for the next Chapter. The octamerization of the λ repres´vet et al., 1999, see also the papers by sor by looping was found by B. Re I. B. Dodd et al., 2001 and 2004. Finally, the paper by J. M. G. Vilar and S. Leibler, 2003, extends the work by Ackers and Arkin to account for the energetics and fluctuations of looping. Papers touching on further aspects of the λ switch are by E. Aurell and K. Sneppen, 2002; A. Bakk et al., ´n and M. C. Mackey, 2004. 2004, and M. Santilla Theory of pattern formation. The pioneer in the field is Alan Turing; it is worth reading his original papers and even looking into his unpublished notes which can be found in the library of the King’s college in Cambridge. They are also available on the WWW. A modern classic in the theory of pattern formation is the work by A. Gierer and H. Meinhardt, 1972. An ongoing topic of interest (and controversy) is the robustness of the morphogen gradient in drosophila embryos. See the paper by B. Houchmandzadeh et al., 2002. A recent theory which explains the experiment is by M. Howard and P. R. ten Wolde (2005); they postulate the existence of a yet unknown corepressor. The Min system. The Min system enjoys ongoing interest in modelling; see the papers by M. Howard, K. Kruse, H. Meinhart, N. S. Wingreen and collaborators. The system is nice since a lot can be done experimentally. Note, however, that despite the modelling approaches all being fairly similar in the underlying philosophy, there are still unanswered questions, and the modelling is not unambiguous. Thus: a problem with a lot of achievements, but still not fully solved; see the discussion by M. Howard and K. Kruse, 2005. One question that remains open is, e.g., why at all an oscillation is used to determine the midplane. Other bacteria have developed stable patterns, see M. Howard, 2004. Modern approaches. Min is nice, but drosophila is nicer. Relating back to the problem of transcriptional control on the chromatin level it is of enormous interest and difficulty to study the structure formation in embryo de-

Network Dynamics I: Deterministic


velopment. Modern approaches employ modelling (J. Reinitz et al., 2003), microfluidics (E. M. Luchetta et al., 2005), and gene network engineering (M. Isalan et al., 2005).


Computational Biology: A Statistical Mechanics Perspective

References G. K. Ackers, A. D. Johnson and M. A. Shea, Quantitative model for gene regulation by λ phage repressor, Proc. Natl. Acad. Sci. USA 79, 1129-1133 (1982) A. Arkin, J. Ross and H. H. McAdams, Stochastic Kinetic Analysis of Developmental Pathway Bifurcation in Phage λ-Infected Escherichia coli Cells, Genetics 149, 1633-1648 (1998) E. Aurell and K. Sneppen, Epigenetics as a First Exit Problem, Phys. Rev. Lett. 88, 048101 (2002) A. Bakk, R. Metzler and K. Sneppen, Sensitivity of OR in Phage λ, Biophys. J. 86, 58-66 (2004) I. B. Dodd, A. J. Perkins, D. Tsemitsidis and J. B. Egan, Octamerization of λ CI repressor is needed for effective repression of PRM and efficient switching from lysogeny, Genes & Development 15, 3013-3022 (2001) I. B. Dodd, K. E. Shearwin, A. J. Perkins, T. Burr, A. Hochschild and J. B. Egan, Cooperativity in long-range gene regulation by the λ CI repressor, Genes & Development 18, 344-354 (2004) J. L. England and J. Cardy, Morphogen gradient from a noisy source, Phys. Rev. Lett. 94, 078101 (2005) C. P. Fall, E. S. Marland, J. M. Wagner and J. J. Tyson, Computational Cell Biology, Springer (2002) A. Goldbeter, Biochemical Oscillations and Cellular Rhythms, Cambridge University Press (1996) A. Gierer and H. Meinhardt, A Theory of Biological Pattern Formation, Kybernetik 12, 30-39 (1972) J. B. Gurdon and P. Y. Bourillot, Morphogen gradient interpretation, Nature 413, 797-803 (2001) J. Hasty, J. Pradines, M. Dolnik and J. J. Collins, Noise-based switches and amplifiers for gene expression, Proc. Natl. Acad. Sci. USA 97, 2075-2080 (2000) B. Houchmandzadeh, E. Wieschaus and S. Leibler, Establishment of developmental precision and proportions in the early Drosophila embryo, Nature

Network Dynamics I: Deterministic


415, 798-802 M. Howard, A. D. Rutenberg and S. de Vet, Dynamic Compartmentalization of Bacteria: Accurate Cell Division in E. Coli, Phys. Rev. Lett. 87, 278102 (2001) M. Howard and K. Kruse, Cellular organization by self-organization: mechanims and models for Min protein dynamics, J. Cell Biol. 168, 533-536 (2005) M. Howard, A Mechanism for Polar Protein Localization in Bacteria, J. Mol. Biol. 335, 655-663 (2004) M. Howard and P. Rein ten Wolde, Finding the Center Reliably: Robust Patterns of Developmental Gene Expression, Phys. Rev. Lett. 95, 208103 (2005) K. Huang, Y. Meir and N. S. Wingreen, Dynamic structures in Escherichia coli: Spontaneous formation of MinE rings and MinD polar zones, Proc. Natl. Acad. Sci. USA 100, 12724-12728 (2003) M. Isalan, C. Lermerle and L. Serrano, Engineering Gene Networks to Emulate Drosophila Embryonic Pattern Formation, PLOS Biology 3, e64 (2005) E. M. Luchetta, J. H. Lee, L. A. Fu, N. H. Patel and R. F. Ismagilov, Dynamics of Drosophila embryonic patterning network perturbed in space and time using microfluidics, Nature 434, 1134-1138 (2005) G. Meacci and K. Kruse, Min-oscillations in Escherichia coli induced by interactions of membrane-bound proteins, Phys. Biol. 2, 89-97 (2005) H. Meinhardt and P. A. J. de Boer, Pattern Formation in Escherichia coli: A model for the pole-to-pole oscillations of Min proteins and the localization of the division site, Proc. Natl. Acad. Sci. USA 98, 14202-14207 (2001) H. Merlitz, K. Rippe, K. v. Klenin and J. Langowski, Looping Dynamics of Linear DNA Molecules and the Effect of DNA Curvature: A Study by Brownian Dynamics Simulation, Biophy. J. 74, 773-779 (1998) M. Ptashne, A Genetic Switch. Third Edition: Phage Lambda Revisited, Cold Spring Harbor Laboratory Press (2004) J. Reinitz, S. Hou and D. H. Sharp, Transcriptional Control in Drosophila, ComPlexUs 1, 54-64 (2003)


Computational Biology: A Statistical Mechanics Perspective

B. R´evet, B. von Wilcken-Bergmann, H. Bessert, A. Barker and B. M¨ ullerHill, Four dimers of λ repressor bound to two suitably spaced pairs of λ operators form octamers and DNA loops over large distances, Curr. Biology 9, 151-154 (1999) M. Santill´ an and M. C. Mackey, Why the Lysogenic State of Phage λ Is So Stable: A Mathematical Modeling Approach, Biophys. J. 86, 75-84 (2004) J. J. Tyson, K. Chen and B. Novak, Network dynamics and cell physiology, Nat. Rev. Mol. Cell Biol. 2, 908-916 (2001) J. M. G. Vilar and S. Leibler, DNA looping and physical constraints on transcription regulation, J. Mol. Biol. 331, 981-989 (2003)

Chapter 2 Network Dynamics II: Fluctuations

Fluctuations are natural features of biological networks given that molecules are discrete objects and, in many circumstances in biology, only present in small numbers. To be more specific, what we want to address here are some aspects of the functional role fluctuations can play in biological networks. We do this by way of example, since a general theory of fluctuations in biological systems does not yet exist.1 The examples are chosen as applications of two of the theoretical approaches discussed in the first part of the book, the Langevin-approach and the chemical master equation. The last part of this Chapter is devoted to a more general discussion of fluctuations, coming back to the distinction between intrinsic and extrinsic noise sources, this time in the context of biology.


Noise in signalling

We begin the discussion in this section by considering the dynamics associated with that of a binding site for signalling molecules, say a promoter region for a transcription factor, present at a concentration c. In Chapter 1, we had considered the binding of transcription factors at the operator sites within a model based on differential equations, in which the descriptors were the molecular concentrations. Here, we want to look a bit closer now, and understand the binding and unbinding of the transcription factors as a process leading to fluctuations in the occupancy of the binding sites. How can we characterize these fluctuations? We address this issue first again in the framework of differential equations. The occupancy of the promoter site is, within a deterministic modeling setup, given by a kinetic equation for the time-dependent occupancy of the binding sites n(t) for a given total concentration c of binding molecules, (W. Bialek

1 Is

such a thing to be expected?



Computational Biology: A Statistical Mechanics Perspective

and S. Setayeshgar, 2005) n˙ = k+ c (1 − n(t)) − k− n(t) .


At equilibrium, the binding is determined by detailed balance and associated with the binding free energy   k+ c F = kB T ln . (2.2) k− We now assume that thermal fluctuations affect the rate constants. We linearize eq.(2.1) around the mean value n to obtain ˙ = −(k+ c + k− )δn + c(1 − n)δk+ − nδk− , δn


and, in addition, δk+ δk− δF . − = k+ k− kB T


Together, this yields the expression ˙ = −(k+ c + k− )δn + k+ c(1 − n) δF . δn kB T


We can Fourier transform the equation for δn(t) and obtain the result (Exercise!)   δn(ω) 1 k+ c(1 − n) = . (2.6) δF (ω) kB T −iω + (k+ c + k− )

The total variance of the fluctuations can be obtained by invoking the dissipation-fluctuation theorem2 Z dω h(δn)2 i = Sn (ω) (2.7) 2π where Sn (ω) =

2k+ c(1 − n) , ω 2 + (k+ c + k− )2


which yields h(δn)2 i = n(1 − n) = 2 The


k+ c (1 − n) . k+ c + kc

DFT in this context can be written as Z dω −iω(t−t0 ) hn(t)n(t0 )i = e Sn (ω) 2π


Network Dynamics II: Fluctuations


Note that the inverse of the sum of the rates τc ≡

1 k+ c + k−


is the correlation time of the fluctuations. We now want to generalize this result to the case where the concentration c of transcription factors is allowed to vary spatially, i.e., it obeys a diffusion equation with a sink term, ∂t c(x, t) = D∆c(x, t) − δ(x − x0 )n(t) ˙ . One finds as a result the generalization of eq.(2.6),   δn(ω) k+ c(1 − n) 1 = δF (ω) kB T −iω(1 + Σ(ω)) + τc−1 )



with Z Σ(ω) = k+ (1 − n)

1 d3 k . (2π)3 [−iω + Dk 2 ]


Task. Go through the steps of the above calculation. Discuss the properties of the Fourier-space integral; it is helpful to go back to the discussion of fluctuations in the Ising model in Part I, Chapter 1. Introduce a cutoff in k-space Λ = π/a. The cutoff length a is a microscopic length which we identify with the size of the binding region. If we look for the low-frequency (long-time) limit of the function Σ(ω), we obtain Σ(ω  D/a2 ) ≈ Σ(0) =

k+ (1 − n) , 2πDa


and we find for the spectral density of the occupancy fluctuations the result Sn (ω) ≈ 2k+ c(1 − n)

1 + Σ(0) . ω 2 (1 + Σ(0))2 + τc−2


The allowance for spatial fluctuations in the transcription factor concentration affects the correlation time, since now τcspatial =

1 − n h(δn)2 i + . k− 2πDac



Computational Biology: A Statistical Mechanics Perspective

The diffusion contribution adds a minimum noise level given by the rootmean-square bound per time interval ∆t δnrms >

n(1 − n) . πDac∆t


Let’s now try to see whether this bound makes sense and plug in some numbers. In a bacterial cell, typically Ntf ≈ 100 transcription factors are present in a cellular volume of 1 µm3 ; their diffusion coefficient is D ∼ 3 µm2 /s. With the size of a promoter site of a ≈ 3 nm, one finds πDc ≈ 3/s. Thus the fluctuations in the site occupancy in a time interval ∆t are given by δn > 0.55(1 − n) n

100 Ntf


s . ∆t


Experiments on E. coli show that about 10 % efficiency of gene expression control can be achieved at small n. In order to be consistent with the above limit, the promoter site must have an occupancy time on the order of one minute.


Stochastic cascades

In the previous section we studied how the interaction of a transcription factor with its binding site is affected by noise. In this section, we turn to the problem of how the noise that is affecting one component propagates in a network of several components. This occurs in transcriptional networks, i.e., the chain of biomolecules composed of transcription factors, DNA binding sites, mRNA products and proteins. But this is equally true also in signalling or metabolic networks, where we speak about pathways of biochemical reactions, where at each level fluctuation effects can intervene. One might therefore simply expect that adding level upon level of noise on top of each other might always lead to a completely washed-out output signal. So the question arises what conditions must be satisfied such that a graded input signal can lead to an unambiguous, e.g., all or none, output signal? We want to address this question for a simple model for a signalling cascade which was proposed by M. Thattai and A. van Oudenaarden, 2002. While motivated by real biological cascades, such as phototransduction or protein kinase cascades, the model is not tuned to a direct comparison with such systems - which would require to build in much more detail - but rather

Network Dynamics II: Fluctuations


to get a general idea about the noise robustness of reaction cascades. A recent application of these modelling ideas in a study of noise propagation in gene networks has been performed by J. M. Pedraza and A. van Oudenaarden, 2005. The situation we have in mind is illustrated in Figure 2.1: The cascade is

FIGURE 2.1: Generic model of a biochemical cascade, after M. Thattai and A. van Oudenaarden, 2002. For the explanation of symbols, see text.

built up from a set of species of number Yi , i = 0, ..., N which are linked by differential amplication factors Ci which determine the response of Yi+1 to a change in Yi . The input signal at Y0 is read out at YN . At each level, the concentrations of the species are subject to a noise qi . The model for this cascade that we will formulate is linear to begin with, and hence can be treated analytically using the methods of Part I, Chapter 2. Before we write down the mathematical model for the cascade, we begin with some preliminary considerations that we will later use in the course of the discussion; they also serve as a reminder of how to deal with Langevin equations. For the discussion, we need generic Langevin equations of the form X˙ = g(X) + η(t)


where the noise η(t) is defined by the moments hηi = 0 ,

hη(t)η(t + τ )i = qδ(τ ) .


In these equations, we understand as X the number of a chemical species in a cell. In the simplest case, we can assume that X is produced with a rate k+ and destroyed with a rate k− , where both processes are assumed Poissonian. We then have ˙ X(t) = k+ − k− + η(t) (2.21)


Computational Biology: A Statistical Mechanics Perspective

and we can compute the change in X in a time interval ∆t as hδXi = N+ − N− = (k+ − k− )∆t.


Why did we assume production and destruction of X as Poissonian? In order 2 to make use of the relationship hN± i = hN± i which holds in this case. Thus we have hδX 2 i = (k+ + k− )∆t , (2.23) on the one hand, and, using eq.(2.20) on the other we have 2





dtdt0 hη(t)η(t0 )i = q∆t

hδX i = 0



such that we can identify q = k+ + k− .


We now consider a protein product Y subject to a decay rate γ. We assume that it is produced proportional to an average number b per mRNA transcript, such that the equation reads as Y˙ = kb − γY + η(t) .


In a steady-state, the variance is given by hδY 2 i = (1 + b)hY i


in which the factor b takes into account that the generally occurring ‘bursts’ in protein production will increase the variance above the Poisson level. From eq.(2.26) we find after expansion Y → Y0 + δY for δX the equation for the fluctuations δ Y˙ + γδY = η(t) . (2.28) By Fourier transforming we obtain (iω + γ)δγ(ω) = η(ω) , and thus in time h|δY (ω)|2 i = and finally hδY 2 i =


dω −∞

q ω2 + γ 2

q q = . ω2 + γ 2 2γ

(2.29) (2.30)


Network Dynamics II: Fluctuations


Since hY i = kb/γ, we obtain q = 2kb(1 + b). After these preliminaries we can finally go on to the cascade model. Its Langevin equation reads Y˙ i + γYi = Fi−1 (Yi−1 )


where the decay rates are all assumed equal, and we simply put them equal to one. For the fluctuations we find δ Y˙ i + δYi = Ci−1 δYi−1 + ηi


where the noise variance is hηi (t)ηi (t + τ )i = qi δ(τ ) .


Employing the Fourier representation, we can write hδYi2 (ω)i =

2 2 qi + Ci−1 hδYi−1 (ω)i . 2 1+ω


This equation is of the form ζi = αi + βi−1 ζi−1 where αi ≡

qi Ci , βi ≡ , ζi ≡ hδYi2 (ω)i . 1 + ω2 1 + ω2



This recursion relation can be fully spelled out as ζN = αN + βN −1 αN −1 + βN −1 βN −2 αN −2 + ... + βN −1 · · · β0 α0 .


Let us look for the upper bound of the variances in the limit of an infinite cascade, i.e., for N → ∞. Introducing q ≡ max qi , C ≡ max |Ci |, α ≡ q/(1 + ω 2 ) β ≡ C/(1 + ω 2 ), we have ζN ≤ α(1 + β + ... + β N −1 ) + βζ0 from which the condition ζ∞ ≤

α , 1−β

(2.39) (2.40)

and consequently 2 hδY∞ i≤


dω 1 q ≤ √ 2π 1 + ω 2 − C 2 2 1 − C2



Computational Biology: A Statistical Mechanics Perspective

follows. The result of this calculation, eq.(2.41), shows that the output fluctuations will stay bounded provided that |C| ≤ 1, but they can be larger than the input √fluctuations at any single cascade stage, due to the presence of the factor 1/ 1 − C 2 . But, surprisingly, we will now see how this result can be used to attenuate the noise. For |C| < 1 the output noise will be independent of the input noise, if only the cascade is sufficiently long. To demonstrate this, we again consider a cascade of species Yi with low noise strength qi = q for all i = 1, .., N which is subjected to a high noise at input, q0 > q. The inverse Fourier transform of the recursion relation gives hδYN2 i =


Z qN −m


   2m N X dω qN −m 2m C 2m C = . m 2π (1 + ω 2 )m+1 2 2 m=0 (2.42)

The variance of the output signal at stage N of the cascade contains contributions from the cascade stage itself, and from the input √ carried along. With Stirling’s approximation for the factorial m! ≈ (m/e)m 2πm this reads as hδYN2 i

q = 2

N −1 X

C 2m √ 1+ mπ m=1


q0 + 2

C 2N √ Nπ



While the first term in this expression increases with cascade length N , the second term is attenuated exponentially, since N

e − N0 ∼ √ N


N0 ≡ (− ln(C 2 )) −1


hδYN2 iinput where

is a sort of ‘attenuation length scale’. It can serve as an estimate for the length of the cascade required to beat the noise on the input signal. So much for a simple, linear model. The next step for us now is to see how fluctuation effects can work in systems which interact in a nonlinear fashion. We will discuss this for a - still fairly simple - feedback system in the next section.

Network Dynamics II: Fluctuations



Stochastic focusing

In this section, we discuss the stochastic dynamics of an intracellular negative feedback model system (J. Paulsson and M. Ehrenberg, 2000). In this model, two species X and S are present that regulate each other’s synthesis. The biological motivation is taken from bacterial genomes, the so-called plasmids. Plasmids are self-replicating genomes that are able to self-control their copy number. We want to see how a simple system of two components can achieve this. The macroscopic dynamic equations describing the feedback system read x˙ =

kx −x 1 + αs


s˙ = ks x − kd s . As before, the small letters x and s denote concentration variables. The interpretation of the equations is as follows: the X-molecules multiply autocatalytically (like the plasmids in the motivating example), while the S molecules inhibit X-synthesis by what is called hyperbolic inhibition. This is reflected in the factor 1 (2.47) q≡ 1 + αs in the equation for x: the presence of a high concentration of s inhibits the production of x. Hyperbolic inhibition is a ubiquitous control mechanism arising in various reaction schemes (see J. Paulsson et al., 2000). In the system given by eq.(2.46), the parameters k and α set two characteristic concentration scales, and kd determines how the steady state is approached: for small kd , the approach is oscillatory, while for large values, s remains ‘slaved’ to x. Normalisation of the equations with respect to the steady state values via xr ≡ x/¯ x and sr ≡ s/¯ s leads to x˙ r =

(k − 1)(1 − sr ) xr , 1 + (k − 1)sr


s˙ r = kd (xr − sr ) , so that one immediately sees that in the limit k  1 the rate of synthesis of X behaves has   1 x˙ r ≈ − 1 xr . (2.49) sr


Computational Biology: A Statistical Mechanics Perspective

Thus, high concentrations in S favour the decrease in X, and low concentrations in S favour the increase in X. The system cannot run away to high values of X, though, since the second equation couples the evolution of X directly to that of S. We now want to look into what will happen to the fluctuations in X and S. For this we set up the chemical master equation which corresponds to the reactions between the S and X molecules. Suppose we start with m Xmolecules and n S-molecules. The synthesis of X is governed by the rate gmn = km/(1 + αn). With the step operator E (see Part I, Chapter 2) one obtains the master equation in the form p˙mn = (Em−1 − 1)gmn pmn + (Em − 1)m pmn + ks m(En−1 − 1) pmn + kd (En − 1)n pmn + pmn

(2.50) ∞ X



for m > 0, n ≥ 0; the last term arises from conditioning the distribution on m > 0. This choice reflects the fact that the state with X = S = 0 is an absorbing state. X can be considered a molecule whose absence signals cell death. We now simplify this master equation by the additional assumption that the number of S molecules rapidly adjusts to the number of X molecules; it is thus considered a fast variable and can be eliminated from the equations. The dependence on n in eq.(2.50) is thus dropped and the equation simplifies to p˙m = (Em−1 − 1)gm pm + (Em − 1)mpm + p1 pm . (2.51) This equation will now be analyzed for both noise-free and noisy signals, depending on the conditional variation of S. When the conditional S-variation for a given value of X is negligible, gm =

km . 1 + αm(ks /kd )


If k is large (or, αm(ks /kd ) small), gm becomes constant for large m, and the number of X molecules remains Poisson distributed except at low averages hmi. But when the conditional variation on S is not negligible - as a consequence of a noisy signal - eq.(2.52) has to be replaced by gm = km

∞ X pn|m . 1 + αn n=0


The quasistationary conditional probabilities of n S-molecules given m Xmolecules pn|m are Poissonian with average hnim = mks /kd , since all synthesis and degradation events are independent. The number of S-molecules at

Network Dynamics II: Fluctuations


any given time represents the number of X-molecules in a probabilistic sense, and hence S is a ‘noisy’ slave to the slow variable X.

FIGURE 2.2: X-variance as a function of k for hyperbolic inhibition. Reprinted with permission from J. Paulsson and M. Ehrenberg; Copyright (2000) American Physical Society.

Figure 2.2 shows the variance of the X-molecules as a function of k for the two cases, assuming an average of hmi = 10, and hnim = m. One sees that for the second case, the variance in X can be reduced indefinitely, while it saturates for the former. This shows that it is indeed the variations of S that lead to a reduction in the variation of X. This effect has been dubbed stochastic focusing - it is the paradoxical effect that the fluctuations in one component effectively reduce those in another one.


Fluctuating gene network dynamics

We have already seen in the discussion of the simple example of the λ phage how the combination of gene network elements can lead to complex dynamic behaviour. We will now go further and study a prominent example of a gene network, one which has been built artificially: the repressilator (M. B. Elowitz and S. Leibler, 2000). The repressilator is a gene circuit which consists of inhibitory gates: the binding of a transcription factor at a


Computational Biology: A Statistical Mechanics Perspective

binding site blocks the transcription of the corresponding gene; this is similar to the binding of the λ-repressor dimer at OR3 and blocking the access for the RNA polymerase. In this section we select a different level of description than in the previous sections: we do not want to account explicitly for the molecules binding to their binding sites, but rather prefer to treat each gene and its associated transcription and translation machinery as a composite object which we call a gate - in obvious analogy to the terminology common to electronic circuitry. The transcription and translation actions being combined, a gate will ‘fire’ when expression takes place. In order to build a gene network, we essentially need gates with three types of actions: • a gate which ‘fires’ without input, which corresponds to expression at a basal rate r in the expression of a gene; • a gate which is blocked upon binding a transcription factor; • a gate which fires upon input, corresponding to the binding of a transcription factor - the activated case. All cases are illustrated schematically in Figure 2.3.

FIGURE 2.3: Three types of gene gates: a), left: constitutive transcription; b) middle: inhibition; c) right: activation of gene expression.

In order to model these gates we use a concept from theoretical computer science, a calculus called the π-calculus (R. Milner, 1999). We can understand the gate actions as ‘communication’ events: each gate can receive input or send outputs over communication channels. In our case, the messages sent over the channels are simple synchronizations between output and input actions at the two different gates. The occurrence of an interaction (read: communication) requires both partners to have the corresponding input and output channels available.

Network Dynamics II: Fluctuations


In practice, this works as follows. The action of gate a) of Figure 2.3 which transcribes constitutively, will be written formally as null(u) = τc . (tr(u)|null(u))


Note that this line should not be read as an equation but as a procedural description - as computer code. What is written on the right hand side are the computational steps that are performed. The notation means that the gate is considered a process with a parameter u which is its transcription product; there is no input. The notion of process in this context differs from the one commonly used in physics: this process is an object which can perform actions in time, like a running computer program. While the notation of the π-calculus may need some time of accomodation for a statistical physicist, the use of process algebras and calculi for modelling in systems biology is increasing. A big advantage of these approaches is their modularity (or, in technical language, compositionality). This means that the building blocks of the networks are easy to combine and modify, much easier than is the case in the more classic approaches such as differential equations - an advantage that becomes crucial if one wants to model large systems. Coming back to the formal description of the constitutive gate, the actions of this process are defined as follows: • first perform a stochastic delay τc with a rate c. The rates can be determined according to the Gillespie algorithm simulating the underlying master equation of the stochastic dynamics, see Part I, Chapter 2; • the separating dot means ‘then’; • the term in brackets denotes two parallel processes x and y, (x|y). The second process is a copy of the original process since processes are consumed after they have performed their actions; • tr(u) denotes the process which produces u, and has to be defined separately. The process defining the transcription factor tr(u) is written as as tr(u) =! + τδ .0


There are two actions now among which can be chosen: + denotes choice in this context. The first process is an output action (represented by the exclamation mark ‘!’), followed by a copy of the original process, as already discussed before. The second action is a degradation process, denoted by 0, which occurs after a delay τδ , where δ again is a rate.


Computational Biology: A Statistical Mechanics Perspective

In the same style we can now introduce the regulated gates. The gate b) of Figure 2.3 which is blocked by receiving an input (a transcription factor) can be written as neg(v, u) =?v.τη .neg(v, u) + τc .(tr(u)|neg(v, u))


The second part of this expression should be clear: this is just the constitutive part. The first part means that the gate is ready to receive an input in the form of a transcription factor v which is followed by a stochastic delay; then the system returns to its initial state. Finally the gate c) of Figure 2.3, whose expression activity is enhanced upon the action of a factor v is written as neg(v, u) =?v.τη .(tr(u)|pos(v, u)) + τc .(tr(u)|neg(v, u))


The symmetry between the first and the second parts of the statement is now evident; by comparison to gate b), gate c) both produces its product u constitutively, and by activation with factor v.


The wiring diagram of the repressilator.

We now have all the elements to build the circuits. The wiring diagram of the repressilator is shown in Figure 2.4: three neg-gates are coupled. The wiring

Network Dynamics II: Fluctuations


is expressed by the names of the products and hence the code describing the system is given by neg(v, u)|neg(u, w)|neg(w, v)


Figure 2.5 shows a typical simulation run, provided suitable degradation rates are chosen. As mentioned before, the underlying stochastic simulation technique is based on the Gillespie-algorithm, see Part I, Chapter 2. As was found in the experiments, the system can be made to oscillate, and all transcription factor concentrations go through clearly separated and synchronized rising and falling cycles (M. B. Elowitz and S. Leibler, 2000). The π-calculus-based simulations of the repressilator and of the more complex combinatorial gene networks built by C. C. Guet et al., 2001, have been described in R. Blossey et al., 2005. An interesting result of these simulations is that stochastic effects can indeed play a role for the interpretation of the experimental data, which for some of the combinatorial circuits cannot be explained by simple logical analysis. But this is a complex issue, since another way to explain the data might be that the biological system is not completely described by the model.

FIGURE 2.5: Repressilator oscillations. Shown are the output offers on the transcription factor channels.



Computational Biology: A Statistical Mechanics Perspective

Extrinsic vs. intrinsic noise

So far we have dealt with the interplay of noise in systems of few different, or many alike components. We now take a further step to distinguish more precisely the different contributions to noise, and how they can be disentangled in experiment. P. Swain et al., 2002, have introduced an explicit distinction between extrinsic and intrinsic sources of noise in the context of biological systems. We have seen already before, in Part I, Chapter 2, a definition of these notions from the point of stochastic dynamic systems - a clarification is thus needed of the relation between the two concepts with the same name. Let us discuss what Swain et al. refer to with these notions. The difference between intrinsic and extrinsic noise according to the definition by Swain et al. is easy to grasp from the following example. Consider one particular gene and its protein product - let’s call it A - in a population of genetically identical cells. We want to relate this to the time evolution of the protein concentration or number in a single cell. Even if all cells were identical at a given time, the molecular processes involved in the production of the protein (transcription and translation) are not identical in the cells; the production process of protein A will thus contain fluctuations that are intrinsic to the gene of protein A. At the same time, the actors intervening in the production process, like the RNA-polymerase, are themselves gene products and also display cell-to-cell and time-dependent variations. Since their properties are not controlled by the production of protein A, one can consider their effect as being an extrinsic source of noise to the production of protein A. This means that the variations of the extrinsic factors (like the fluctuations in the number of polymerases) arise independently from the intrinsic variations, but clearly influence them. Obviously, extrinsic sources of noise in this sense are abundant in a cell, in which many processes are coupled. The definition by Swain et al. is thus quite different from the one by van Kampen. In van Kampen’s interpretation, systems like the ones from biology we discuss here have only intrinsic noise to begin with, since they are built up from molecular constituents - nowhere are deterministic processes in sight. The extrinsic noise by Swain et al. thus expresses the fact that the system can be separated by way of analysis way into several noise-dependent mecha-

Network Dynamics II: Fluctuations


nisms.3 Let us see where this assumption leads us. We now formulate the mathematical description for the distinction between the two types of noise suggested by P. Swain et al., 2002. We represent the intrinsic and extrinsic properties by two vectors I and E, whose components list the different noise sources. Further, let us identify by Pk the expression level of a specific gene in cell k. From an experimental measurement taken on an ensemble of N cells, the Pk can be deduced and averaged to obtain the moments of the corresponding protein distribution in the form Z N 1 X m Pk ≈ dE dI P m (E, I)p(E, I) N



where p is the joint probability distribution for intrinsic and extrinsic variables, and P m (E , I) is the measured expression level for particular values of E and I. Using the product rule of probability distributions we can separate this into contributions of intrinsic and extrinsic variables by invoking conditional probabilities (see Part I, Chapter 2) Z Z N 1 X m Pk ≈ dE p(E) dI P m (E, I) p(I|E) , N



and we can introduce the definition of the average over intrinsic variables from the second integral Z hP m (E)i ≡ dI P m (E, I) p(I|E) . (2.61) By contrast, averages over extrinsic variables can be indicated with an overline in the form N 1 X m m hP i = Pk . (2.62) N k=1

3 The

doubling of the notions of extrinsic and intrinsic with different meanings, once in the context of general stochastic processes and once in a specific biological context, is unfortunate. In the biological context, a better notion would probably have been to talk about cis-noise and trans-noise, where cis-noise is stochasticity in the same stochastic reaction pathway, and trans-noise is that exerted from one stochastic reaction pathway onto another. Both cis- and trans-noise are intrinsic in character.


Computational Biology: A Statistical Mechanics Perspective

which amounts to an average over both intrinsic and extrinsic noise sources. If we want to quantify the noise strength by standard deviations over mean, we write for the total noise, as measured in experiment 2 ηtot


1 N



Pk2 − P 1 N


1 N



k 2


2 ,


which is equivalent to  2 hP 2 i − hP i = .  2 hP i


 2 hP i2 − hP i hP 2 i − hP i2 2 2 =  = ηint + ηext . 2 +  2 hP i hP i


2 ηtot

This can be rewritten as 2 ηtot

In order to distinguish experimentally between these two different contributions, one has to measure hP i2 . This quantity can be obtained as follows. If two identical copies of a particular gene were present in the same cell k, their (1) (2) expression levels were Pk and Pk . These would have the same extrinsic sources of noise, but different intrinsic variables. Consequently, Z Z N 1 X (1) (2) Pk Pk ≈ dE dI1 dI2 P (E, I1 )P (E, I2 )p(I1 , I2 , E) N k=1 Z 2 Z = dEp(E) dIP (E, I)p(I|E) = hP i2 .


The verification of this scenario can be made by fluorescent tagging of two genes controlled by identical regulatory sequences. These experiments have been performed by M. B. Elowitz et al., 2002, in E. coli. In the absence of intrinsic noise, the concentrations of both proteins, as measured by their fluorescence, would fluctuate in a correlated way. In a population of cells this then gives rise to a certain distribution of protein levels due to extrinsic noise alone. In the presence of intrinsic noise, the correlations between the two protein concentrations in each single cell decrease, reflecting different expression levels for the proteins.

Network Dynamics II: Fluctuations


Figure 2.6 shows the experimental results in the form of a plot of noise levels as a function of a measure for population-averaged transcription rates, distinguishing between extrinsic and intrinsic noise contributions.

FIGURE 2.6: Cellular genetic noise in E. col i . Total, intrinsic and extrinsic noise as a function of fluorescence levels (population means) characterizing transcription rates. Reprinted with permission from Science from M. B. Elowitz et al., 2002.


Computational Biology: A Statistical Mechanics Perspective

Additional Notes Fluctuations, noise and robustness (see next Section) are becoming topics of increasing interest in the computational biology community. This is to a large part, but not exclusively, due to the influx of statistical physicists into the field. An early contribution is by H. H. McAdams and A. Arkin, 1997. The effects of noise in the standard model of gene expression (DNA → RNA → protein) has been modelled and discussed with respect to experimental work in particular by J. Paulsson (2004, 2005). Similar work is due to M. Thattai and A. van Oudenaarden, 2001, and by A. Becskei, B. B. Kaufmann and A. van Oudenaarden, 2005. Pi-calculus and process algebras. There is a whole literature developing on systems biology modelling based on approaches by theoretical computer scientists, in particular by specialists on programming languages. The seminal work is by A. Regev, 2002. Intrinsic vs. extrinsic noise. This topic enjoys growing interest in the bio-community. A recent application of this concept is by A. ColmanLerner et al., 2005, to the cell-fate decision system. Parameter problems. A very general modelling problem in systems biology is the lack of knowledge of parameters - a problem of which already A. Turing was well aware. A paper on statistical mechanics approaches to systems with many poorly known parameters is by K. S. Brown and J. P. Sethna, 2003.

Network Dynamics II: Fluctuations


References A. Becskei, B. B. Kaufmann and A. van Oudenaarden, Contributions of low molecule number and chromosomal positioning to stochastic gene expression, Nat. Genet. 37, 937-944 (2005) W. Bialek and S. Setayeshgar, Physical limits to biochemical signalling, Proc. Natl. Acad. Sci. USA 102, 10040-10045 (2005) R. Blossey, L. Cardelli and A. Phillips, A compositional approach to the stochastic dynamics of gene networks, Transactions in Computational Systems Biology IV, 99-122 (2006) K. S. Brown and J. P. Sethna, Statistical mechanical approaches to models with many poorly known parameters, Phys. Rev. E 68, 021904 (2003) A. Colman-Lerner A. Gordon, E. Serra, T. Chin, O. Resnekov, D. Endy, C. G. Pesce and R. Brent, Regulated cell-to-cell variation in a cell-fate decision system, Nature 437, 699-706 (2005) M. B. Elowitz and S. Leibler, Synthetic gene oscillatory network of transcriptional regulators, Nature 403, 335-338 (2000) M. B. Elowitz, A. J. Levine, E. D. Siggia and P. S. Swain, Stochastic Gene Expression in a Single Cell, Science 297, 1183-1186 (2002) C. C. Guet, M. B. Elowitz, W. Hsing, S. Leibler, Combinatorial Synthesis of Genetic Networks, Science 296, 1466-1470 (2002) H. H. McAdams and A. Arkin, Stochastic mechanisms in gene expression, Proc. Natl. Acad. Sci. 94, 814-819 (1997) R. Milner, The π-calculus, Cambridge University Press (1999) J. Paulsson, O. Berg and M. Ehrenberg, Stochastic focusing: Fluctuationenhanced sensitivity of intracellular regulation, Proc. Natl. Acad. Sci. USA 97, 7148-7153 (2000) J. Paulsson and M. Ehrenberg, Random Signal Fluctuations Can Reduce Random Fluctuations in Regulated Components of Chemical Regulatory Networks, Phys. Rev. Lett. 84, 5447-4450 (2000) J. Paulsson, Summing up the noise in gene networks, Nature 427, 415-418 (2004)


Computational Biology: A Statistical Mechanics Perspective

J. Paulsson, Models of stochastic gene expression, Physics of Life Reviews 2, 157-175 (2005) J. M. Pedraza and A. van Oudenaarden, Noise propagation in Gene Networks, Nature 307, 1965-1969 (2005) A. Regev, Computational Systems Biology: A Calculus for Biomolecular Knowledge, PhD-Thesis University of Tel Aviv (2002) P. S. Swain, M. B. Elowitz and E. D. Siggia, Intrinsic and extrinsic contributions to stochasticity in gene expression, Proc. Natl. Acad. Sci. USA 99, 12795-12800 (2002) M. Thattai and A. van Oudenaarden, Intrinsic noise in gene regulatory networks, Proc. Natl. Acad. Sci. USA 98, 8614-8619 (2001) M. Thattai and A. van Oudenaarden, Attenuation of Noise in Ultrasensitive Signaling Cascades, Biophys. J. 82, 2943-2950 (2002)

Chapter 3 Networks: Structure


Networks as graphs

After our study of biological networks on a small scale - i.e., network components - we now ask how a whole biological network can be defined and characterized. The starting point of our discussion of network structure is the ¨ s and Renyi, 1959, random network, or random graph introduced by Erdo which will serve us as a convenient null model. The ER-graph consists of n nodes or vertices, joined by links or edges between the vertices which are placed at random with independent probability p; an illustration is given in Figure 3.1. The construction allows to introduce the notion of an ensemble of graphs of n vertices, G(n, p), in which each graph is present with the probability corresponding to the number of its edges.

FIGURE 3.1: Random network, or Erd¨os-Renyi graph a). The graph under b) is a represents a scale-free network, to be discussed in the text below. Reprinted with permission from Macmillan Publishers Ltd., R. Albert et al., 2000.)



Computational Biology: A Statistical Mechanics Perspective

The random graph has the following basic properties: • The degree of a vertex is the number of edges connected to it. The average degree of a graph is then given by z =2×

n(n − 1)p = (n − 1)p 2n


where the factor of two takes into account that each edge is connected to two vertices. Hence there are no edges going out and ending at the same vertex. • The probability pk that a vertex in an Erd¨os-Renyi graph has degree k is given by the binomial distribution  pk =

n−1 k

pk (1 − p)n−1−k


which in the limit n  kz yields a Poisson distribution pk =

z k −z e . k!


The degree distribution is thus strongly peaked around the mean z, and decays rapidly for large degrees. Exercise. Show that ER-graphs are always treelike. A graph is called treelike if each of its cycles (or circuits: say internal loops) has at most one common vertex. The question can be posed whether the random graph model applies to biological systems (or to other existing network-like systems, for that matter). An attempt in this direction was made by H. Jeong et al., 2000, in the context of a study of metabolic networks of 43 different organisms. A metabolic network is the network of biochemical compounds whose transformations are needed to sustain body function. It was generally found that these large-scale networks display asymmetric degree distributions with rather ‘fat’ tails, and could be characterized by algebraic behaviour for a range of degrees, k. In analogy to critical phenomena - think of the algebraic behaviour of correlation functions near a critical point discussed in Part I, Chapter 1, such networks have been dubbed ‘scale-free’. In subsequent work, H. Jeong et al., 2001, considered the interaction network built from proteins in yeast Saccharomyces cerevisiae, consisting of 1870 edges and 2240 identified physical interactions among the proteins. Their

Networks: Structure


finding is shown in Color Figure 8. Again, the degree distribution displays a fat tail, and the authors fit it to the expression p(k) =

A exp(−k/kc ) (k + kD ) γ


where kD ≈ 1 and kc ≈ 20 are short- and large scale cutoffs. Evidently, the presence of cutoffs shows that the network is not strictly scale-free. This notion has indeed been much abused in the literature, and we will come back to it below in more detail. At present, it is important to also add that the algebraic behaviour is observed only over a quite limited range of k-values, sometimes hardly more than one order of magnitude. The exponent γ is quite typically found to line in the range 2 < γ < 3. Independently of the issue whether the graph is truly scale-free or not, already the appearance of a fat tail in the degree distribution obviously takes the protein interaction network out of the basic random graph paradigm. As a consequence, in order to understand these networks we need new theory. Some additional observations provide further indication what this new theory should be able to deliver. The protein network is an example of a network which is undirectional, i.e. the interaction of proteins does not distinguish between the partners. This does not hold for all networks: e.g., we may want to study a network of transcriptional interactions. In such a network, interactions are directional. A transcription factor will activate a specific gene (we have seen in Chapter 1 how this works), such that we need to have edges represent not just a basic ‘interacts with’, but also edges which are directed from one node to another. A typical subgraph is indicated schematically in Figure 3.2.


An undirected and a directed graph.


Computational Biology: A Statistical Mechanics Perspective

The fact that the construction of networks as graphs must reflect the particular identity of the interacting components explains immediately why the knowledge of degree distributions is not enough to characterize them. For directed networks, we have to distinguish between ingoing and outgoing degrees; but there are probably many deeper levels of structure as well. Also, it may not be sufficient to look at only one type of network (it may be selected only because of the availability of data): S. Maslov and K. Sneppen, 2002, have shown that the connectivity of protein interaction networks and transcriptional regulation networks in the yeast Saccharomyces cerevisiae show important correlations. The size of the networks used in this study amount to 4549 physical interactions (edges) between 3278 proteins (vertices) for the protein network, and 1289 positive or negative direct transcriptional interactions for 682 proteins. In order to go beyond the ER-graph-network paradigm, we first list a number of important network characteristics and then show how to compute them from a given degree distribution by the method of generating functions. We then move on to a more general statistical mechanical theory of networks.


Probability generating functions and network characteristics

Probability generating functions. We have encountered generating functions in previous Chapters, but only in this application they will rise to the height of their powers. We begin by taking the vertex degree distribution of the random network introduced before, pk , and define, following (M. E. J. Newman, 2003), the generating function G0 (x) =

∞ X

pk xk .



The vertex degree distribution pk can be generated from G0 by differentiation, 1 dk G0 pk = . (3.6) k! dxk x=0 G0 has the following properties: 1. If G0 is properly normalized, then G0 (1) =

X k

pk = 1.


Networks: Structure


2. The average degree hki is found from computing X dG0 (x) 0 ≡ G (1) = kpk = hki . 0 dx x=1



3. For the n-th moment of pk one has   n X d hk n i = k n pk = x G0 (x) . dx x=1


We may of course also consider generating functions of probability distributions other than the degree distribution. E.g., the probability that a vertex that can be reached along one edge from a starting vertex has (k −1) outgoing edges is given by (Exercise!) pk qk−1 = k P . j jpj


The average degree of the vertex then follows as ∞ X

kqk =


hk 2 i − hki . hki


This is the average degree of a vertex two edge distances away from the starting vertex. Based on this result, we can define the number of second neighbors in the graph by ! X z2 = kqk · hki = hk 2 i − hki . (3.12) k

The probability generating function based on qk finally reads as G1 (x) =

∞ X k=0

qk xk =

G00 (x) z1


where z1 ≡ hki. We can now list further important network measures: • The clustering coefficient C of a random graph is given by " #2  2 1 X z 2 z−1 C≡ kqk = c + nz n v z



where cv is the coefficient of variation of the degree distribution (the ratio of standard deviation to mean).


Computational Biology: A Statistical Mechanics Perspective

• The diameter d of a graph is given by the maximum distance between any two connected vertices in the graph. It can be shown that the fraction of all possible graphs with n vertices and m edges for which d ≥ c ln n for some constant c tends to zero in the limit n → ∞. • A giant component is a subgraph whose size scales linearly with the graph size. It corresponds to the infinite cluster in percolation theory.1 The random graph has a phase transition (akin to the percolation transition on a lattice) in which a giant component is present whenever the Molloy-Reed criterion holds ∞ X

k(k − 2)pk = 0 .



• For networks which contain a giant component the average vertex-vertex distance ` is given by ln(n/z1 ) + 1. (3.16) `= ln(z2 /z1 ) Task. Characterize correlations in the ER-graph. Robustness. The degree distribution allows to characterize networks according to their robustness or resilience to attack. R. Albert et al., 2000, compared the properties of an ER-graph and a scale-free graph upon edge removal. These networks differ in a characteristic way: since vertices and edges in ER-graphs are more ‘alike’, its properties such as the average degree change fairly little when edges are removed randomly or even on purpose, i.e., when vertices with a high degree are removed preferentially, simulating an attack on the network. By contrast, the removal of few, highly connected nodes in a scale-free network affects its structure very much. Exercise. Study the robustness of the network described by the generating function 1 G0 (x) = ((n − 1)x3 + xn ) . (3.17) n Picture this network. How does its diameter change as a function of n?

1 Percolation

is a geometric phase transition: consider the case of a two-dimensional regular lattice with all sites occupied. Now one places bonds between the sites with a given probability. A question one can pose now is at what concentration of occupied sites will a cluster arise which spans the whole lattice: this is the infinite cluster. Its appearance (the percolation transition) and the statistical properties of this cluster share many features of thermal phase transitions, but there is of course no free energy in this case.

Networks: Structure


We are now ready for a more systematic approach, along the lines of equilibrium statistical mechanics of the first Chapter of Part I. As there, the notion of entropy proves to be highly useful.


Statistical mechanics of networks

A statistical mechanical theory of networks can be developed for exponential random graphs (J. Park and M. E. J. Newman, 2004). If G is an element of a set of graphs G, we can obtain its probability distribution P (G) by maximizing the Gibbs entropy X


P (G) ln P (G)



P P under the constraints G P (G)xi (G) = hxi i and G P (G) = 1, where xi , i = 1, ..., r are a set of graph observables. The derivation of P (G) can now be performed in complete analogy to our discussion of the thermal ensembles in Part I (Task!), leading to the result P (G) =




θi xi (G)



e−H(G) . Z


As in the thermal case, the expression eq.(3.19) involves a set of Lagrange multipliers θi . Examples. The simplest example is the random graph with a fixed number of vertices n. If we want to characterize the graph only by the mean number of edges, hmi, the Hamiltonian in eq.(3.19) is chosen to be H(G) = θm(G)


We can now evaluate the partition function Z for an ensemble of simple undirected graphs on n vertices. An (n × n) adjacency matrix can be defined by  1 i ↔ j, σij = (3.21) 0 else where the double-arrow symbol ↔ denotes i connected to j. With the adjacency matrix we can represent the number of edges as m=

X i 1/2, β > 0 can be studied in more detail by the rate equation for the evolution of the number of vertices of degree k, Pk (n) when the network as a whole has n vertices. The degree of a node increases by one at a rate ak = β +(1−δ)k. We can then write down a rate equation of the form 1 dPk (n) = (ak−1 Pk−1 − ak Pk ) + Gk dn n


where the first two terms account for the increase of a vertex degree by one. The last term is a source term for the introduction of new vertices with k edges,

Networks: Structure


with a of the edges created by duplication and b = k − a by diversification. The probability of the duplication process is X s gdup = xs (1 − δ)a δ s−a (3.37) a s≥a

where ps = ns /n is the probability of a vertex of degree s chosen for the duplication process. The probability of diversification is gdiv = β b

e−β b!

such that the full expression for Gk is given by X X s e−β Gk = xs (1 − δ)a δ s−a β b . a b!



a+b=k s=a

Since the nk grow linearly in n, we can plug the relation Pk (n) = npk into the rate equation and obtain     1+β β Gk k+ pk = k − 1 + pk−1 + , (3.40) 1−δ 1−δ 1−δ which looks very much like a recursion relation for the pk - but it isn’t, since Gk depends on all ps for s ≥ k. It can, however, be turned into a recursion relation in the limit k → ∞. In this limit, the main contribution to Gk arises for small values of b, and the summand is sharply peaked around s ≈ k/(1 − δ). We may then replace the lower limit by s = k and ps by its value at s = k/(1 − δ). Further, anticipating that pk decays as k −γ we introduce the ansatz ps = (1 − δ)γ pk and finally simplify Gk according to Gk ≈ (1 − δ)γ pk

X s=k


  ∞ X e−β s = (1 − δ)γ−1 pk (1 − δ)k δ s−k βb k b!



since the binomial sum equals (1 − δ)−1 . Summing this all up, in the limit k → ∞, pk is found to have a power-law behaviour pk ∼ k −γ with the value of γ fixed by the relation γ(δ) = 1 +

1 1 − 1−δ (1 − δ)2−γ


We thus find that γ(δ) sensitively depends on δ, but not on β. Choosing a value of δ = 0.53, as can be suggested from observations (A. Wagner, 2001), a value of γ = 2.373.. is found, in accord with what we had indicated before.


Computational Biology: A Statistical Mechanics Perspective

A particularly important point to notice is that this result is an asymptotic one. The convergence to the asymptotic limit in which the power laws reign is very slow. Even for vertex numbers of n = 106 , the power-law regime is only reached over two orders of magnitude of degrees k. Compare this with the data for real protein-networks...

Networks: Structure


Additional Notes The literature on networks emerging within the statistical physics community has literally exploded. An early review was written by R. Albert and ´si, 2002. The mathematical theory of random graphs in the A. L. Baraba ¨ s and A. Renyi, 1959, is summarized by tradition of the pioneers P. Erdo ´s, 2000. B. Bolloba Small worlds. Another important network paradigm that has been uncovered in the last years is that of small-world networks (D. J. Watts and S. H. Strogatz, 1998). The authors studied a random rewiring process of a completely regular network. Initially, the model network is placed on a one-dimensional ring and up to k neighbouring points are linked. Then links are broken and rewired at random across the lattice. Upon the introduction of only a few of these ‘shortcuts’, a significant clustering property of the network arises, characterizing an intermediate stage between complete regularity on the one side, and complete randomness on the other. The important feature of the small-world model is that distant points on the network get access to each other: only a small number number of short-cuts will thus allow the network, although being still fairly sparse, to ‘communicate’ efficiently across all nodes. This phenomenon has acquired some fame through the famous story of ‘six degrees of separation’, referring to the number of acquaintances needed to pass between two randomly selected people in the world. M. E. J. Newman at al., 2000, formulated an exactly solvable mean-field model for the small-world network, but despite being ‘just a meanfield solution’, it is already quite demanding. Motifs and modules. The educated reader might wonder why the notion of ‘motifs’ was not discussed in the main text. Well, we make up for that now a little bit. For details, the reader is asked to consult the book by U. Alon, 2006. Biological networks may display particular global features, but they will also be characterized by the properties of local neighbourhoods of vertices. This has been noticed via the occurrence of particular network motifs (R. Milo et al., 2002, S. S. Shen-Orr et al., 2002). Figure 3.3 shows a prominent example, the feedforward loop.2 While in a random graph subgraphs are more likely to be trees, functional subgraphs of networks are more likely to have a higher number of edges than a subgraph in a random network. Again the distinction between directed and undirected 2 Note

that the innocently looking graph to the right in Fig. 3.2 has the same topology!


Computational Biology: A Statistical Mechanics Perspective

graphs matters.

FIGURE 3.3: A feedforward loop. Left: basic structure of the loop; right: an example of a feedback loop in the gene network of E . col i , after S. S. ShenOrr, 2002.

There have been meanwhile a number of studies focussing on network motifs, e.g. for E. coli and S. cerevisiae (E. Yeger-Lotem et al., 2004). In particular the feedforward loop has interesting features which have been recently discussed (S. Mangan and U. Alon, 2003; S. Kalir et al., 2005). Eight different types of feedforward loops have been identified, depending on the actions of the transcription factors (activating or repressing). In E. coli, aside from the feedforward loop, two other motifs have been detected as being statistically overrepresented in the data: the so-called single input module, in which a single transcription factor regulates a set of transcriptional units, the operons, see our discussion of the phage lambda. A third motif is the dense overlapping regulon, in which a number of operons is regulated by a set of transcription factors. A review of the different types of motifs that have been detected is by D. M. Wolf and A. A. Arkin, 2003). The notion of a motif generalizes, on a higher network level, to that of a network module, comprising as separately functioning network elements (E. Ravasz et al., 2002). The distinction between a motif and a module is not evident, as Wolf and Arkin note themselves. The existence of motifs can be used to construct an algorithmic procedure for the detection and significance decision of local network elements. The deterministic motifs are generalized to probabilistic motifs in which edges oc-

Networks: Structure


cur with a certain likelihood. This procedure, called graph alignment due to its analogy to sequence alignment, can be based on a suitably defined scoring ¨ssig, 2004). function, i.e., a Hamiltonian for the motifs, Hscore (Berg and La Finally, it should be stressed that the overrepresentation of certain motifs does not at all guarantee that these are the important building blocks in the system. Biological networks, after all, have evolved over the course of many thousands of years, and the networks hence carry a history along which is very difficult to take into account in the present description of networks (see, e.g., E. Dekel et al., 2005, on the selection of the feedforward motif). As useful as such notions as motifs and modules are as a work guide at present (for a discussion, see the paper by L. H. Hartwell et al., 1999), as little is known about their true biological significance.


Computational Biology: A Statistical Mechanics Perspective

References R. Albert, H. Jeong and A.-L. Barab´asi, Error and attack tolerance of complex networks, Nature 406, 378-380 (2000) R. Albert and A.-L. Barab´asi, Statistical mechanics of complex networks, Rev. Mod. Phys. 74, 47-97 (2002) U. Alon, Introduction to Systems Biology and the Design Principles of Biological Networks, Chapman & Hall/CRC (2006) A.-L. Barab´ asi and R. Albert, Emergence of Scaling in Random Networks, Science 286, 509-512 (1999) J. Berg and M. L¨ assig, Correlated Random Networks, Phys. Rev. Lett. 89, 228701 (2002) J. Berg and M. L¨ assig, Local graph alignment and motif search in biological networks, Proc. Natl. Acad. Sci. USA 101, 14689-14694 (2004) B. Bollob´ as, Random Graphs, Academic (2000) E. Dekel, S. Mangan and U. Alon, Environmental selection of the feedforward loop circuit in gene-regulation networks, Phys. Biol. 2, 81-88 (2005) P. Erd¨ os and A. Renyi, On random graphs, Publ. Math. Debrecen, 6, 290-297 (1959) J.-D. J. Han, D. Dupuy, N. Bertin, M. E. Cusick and M. Vidal, Effect of sampling on topology predictions of protein-protein interaction networks, Nat. Biotech. 23, 839-844 (2005) J. H. Hartwell, J. J. Hopfield, S. Leibler and A. W. Murray, From molecular to modular cell biology, Nature 402, C47-C50 (1999) H. Jeong, B. Tombor, R. Albert, Z. N. Oltvai and A.-L. Barab´asi, The large-scale structure of metabolic networks, Nature 407, 651-654 (2000) H. Jeong, S. P. Mason, A.-L. Barab´asi and Z. N. Oltvai, Lethality and centrality in protein networks, Nature 411, 41-42 (2001) S. Kalir, S. Mangan and U. Alon, A coherent feed-forward loop with a SUM input function prolongs flagella expression in Escherichia coli, Mol. Syst. Biol., msb44100010 (2005)

Networks: Structure


J. Kim, P. L. Krapivsky, B. Kahng and S. Redner, Infinite-Order Percolation and Giant Fluctuations in a Protein Interaction Network, Phys. Rev. E 66, 055101 (2002) S. Mangan and U. Alon, Structure and function of the feed-forward loop network motif, Proc. Natl. Acad. Sci. USA 100, 11980-11985 (2003) S. Maslov and K. Sneppen, Specificity and Stability in Topology of Protein Networks, Science 296, 910-913 (2002) R. Milo, S. S. Shen-Orr, S. Itzkovitz, N. Kashtan, D. Chklovskii and U. Alon, Network Motifs: Simple Building Blocks of Complex Networks, Science 298, 824-827 (2002) M. E. J. Newman, Random graphs as models for networks, in Handbook of Graphs and Networks, S. Bornholdt and H. G. Schuster (eds.), (2003) M. E. J. Newman, C. Moore and D. J. Watts, Mean-field Solution of the Small-World Network Model, Phys. Rev. Lett. 84, 3201-3204 (2000) J. Park and M. E. J. Newman, The statistical mechanics of networks, Phys. Rev. E 70, 055101 (2004) E. Ravasz, A. L. Somera, D. A. Mongru, Z. N. Oltvai and A. L. Barab´asi, Hierarchical Organization of Modularity in Metabolic Networks, Science 297, 1551-1555 (2002) S. S. Shen-Orr, R. Milo, S. Mangan and U. Alon, Network motifs in the transcriptional network of Escherichia coli, Nat. Genet. 31, 64-68 (2002) D. Sprinzak and M. B. Elowitz, Reconstruction of genetic circuits, Nature 438, 443-448 (2005) A. Wagner, The yeast protein interaction network evolves rapidly and contains few redundant duplicate genes, Mol. Biol. Evol. 18, 1283-1292 (2001) D. J. Watts and S. H. Strogatz, Collective dynamics of ‘small-world’ networks, Nature 393, 440-442 (1998) D. M. Wolf and A. P. Arkin, Motifs, modules and games in bacteria, Curr. Op. Microbiol. 6, 125-134 (2003) E. Yeger-Lotem, S. Sattath, N. Kashtan, S. Itzkovitz, R. Milo, R. Y. Pinter, U. Alon and H. Margalit, Network motifs in integrated cellular networks of transcription-regulation and protein-protein interaction, Proc. Natl. Acad.


Computational Biology: A Statistical Mechanics Perspective

Sci. USA 101, 5934-5939 (2004)


acetylation, 168 adiabatic, 5 Affymetrix chips, 117 alpha-helix, 95 amino acids, 88, 92 anti-sense strand, 84 applied work, 54 arc diagram, 89 asymmetric random walk, 52 autocorrelation function, 42 autoregulatory, 187 average degree, 230 average ignorance, 2

comparative information, 145 compositionality, 219 conditional probability, 42 conservation laws, 66 contact matrix, 89 contour surfaces, 7 cooperative effect, 99 cooperative interaction, 176 cooperative phenomenon, 32 cooperativity parameter, 106, 108 copy number, 215 correlation length, 16 covariance, 12 covariance matrix, 12 critical exponent, 24 Crooks fluctuation theorem, 56 cross-hybridization, 117 cumulant expansion, 54 cumulants, 12 Curie’s law, 27

bending energy, 34 Bessel function, 170 beta-sheet, 95 bimodal distribution, 175 binding free energy, 157 biochemical pathways, 210 biochip, 117 biological network, 229 Bjerrum length, 159

Debye screening length, 160 Debye-H¨ uckel equation, 160 denaturation bubbles, 100 denaturation loop, 100 dense overlapping regulon, 242 detailed balance, 51 diffusion equation, 43 diffusion matrix, 49 diffusion term, 57 directed polymer, 73 disorder, 176 dissipated work, 54 dissipation-fluctuation, 50, 61, 208 dissipation-fluctuation theorem, 41 division free energy, 141 DNA, 83 DNA microarray, 117

canonical distribution, 4 canonical ensemble, 3 cDNA, 114 characteristic function, 12 chemical master equation, 66 chemical potential, 4, 8 chemical reactions, 65 chromatin, 167 chromatin remodelling, 168 circle diagram, 92 clustering coefficient, 233 coefficient of variation, 233 collision diagram, 65 communication channel, 218


248 double-stranded RNA, 87 drift term, 57 drift vector, 49 edge, 229 effective potential, 24 Einstein relation, 61 energetic factor, 152 energy, 3 ensemble average, 2, 42 enthalpy, 10 entropic elasticity, 33 entropy, 2 entry-exit angle, 169 equations of state, 8 equilibrium free-energy, 54 equilibrium state, 6 ergodicity, 2 Euler characteristic, 92 Euler equation, 8 excluded-volume effects, 104 exon, 113 exponential random graph, 235 extensive, 5 extreme value distribution, 70 extrinsic noise, 63, 207, 222 feedforward loop, 241 finite-size scaling, 122 first-order homogeneous, 6 first-order phase transition, 25 fixed genus, 138 fixed points, 17 fluctuation theorem, 52 Fokker-Planck equation, 57 fold, 88 forward partition function, 106 Fourier coefficients, 196 free energy, 4 free energy landscape, 139 freely jointed chain, 32, 34 fugacity, 101 funnel, 153 gap exponent, 27

Index gapless alignment, 69 gapped alignment, 72 gate, 218 Gaussian distribution, 11 Gaussian white noise, 61 gene, 113 gene duplication, 237 generator of a flow, 17 genetic networks, 185 genus, 92 geometric fit, 156 giant component, 234 Gibbs distribution, 4 Gibbs free energy, 10 Ginzburg-Landau theory, 23 glass phase, 141 glass transition temperature, 155 global alignment, 70, 73 graded input signal, 210 grand canonical ensemble, 4 grand canonical potential, 10 graph alignment, 243 graph diameter, 234 graph ensemble, 229 green fluorescent protein, 198 Guoy-Chapman length, 161 gyration radius, 104 Hamiltonian, 13 Hartree equation, 136 heat, 5 Helmholtz decomposition, 166 Helmholtz free energy, 10 Hermitian matrix, 134 high-temperature limit, 27 histone proteins, 167 Hopf-instability, 200 hybridize, 84 hydrophobicity, 95 hyperbolic inhibition, 215 hypergeometric function, 138 in vitro, 97 infinite cluster, 234 information, 87

Index information and structure, 87 inhibitory gates, 217 instability, 195 intensive variables, 8 intrinsic noise, 63, 207, 222 intron, 113 Ising model, 13 Itˆ o-Stratonovich dilemma, 63 joint probability density, 42 jump moments, 57, 64 Kinefold, 143 KPZ equation, 74 Kratky-Porod model, 33 Kuhn length, 34, 143 lack of information, 2 Lagrange multiplier, 3 Laguerre polynomial, 138 Langmuir isotherm, 118 Latin square, 118 law of mass action, 67 laws of thermodynamics, 5 Legendre transform, 9 line-charge density, 170 linear phase, 70 linearization, 20 link, 229 linker histone, 167 Liouville equation, 161 logarithmic corrections, 31 logarithmic phase, 70 long-range inhibition, 195 Lorentzian model, 165 low-temperature limit, 142 lower critical dimension, 31 lysogenic phase, 186 lytic phase, 186 macrostate, 2 Manning condensation, 162 Manning parameter, 162 Markov property, 43 master equation, 45

249 maximum average ignorance, 3 McCullogh-Pitts threshold, 174 mean, 12 mean-field approximation, 19 mean-field solution, 19 melting, 97 messenger RNA, 86 metabolic network, 230 method of steepest descent, 21 methylation, 168 microstate, 1, 2 modified Bessel function, 170 molecular beacon, 122 Molloy-Reed criterion, 234 molten phase, 140 moments, 12 morphogen, 194 multivariate, 11 mutation, 150, 237 native-like contact, 153 Needleman-Wunsch, 73 neighbour-dependent coupling, 18 net, 143 network module, 242 network motifs, 241 Neumann boundary condition, 196 Neumann function, 170 node, 229 nonequilibrium quantity, 54 nucleosomal core particle, 167 nucleosome, 167 number of RNA pseudoknots, 136 one-step processes, 48 operon, 186 order of the transitions, 65 order parameter, 22, 103 order parameter field, 20 ordered phase, 27 orientational polarization, 165 Ornstein-Uhlenbeck process, 44 OSF persistence length, 171 partition function, 4

250 periodic boundary condition, 14 persistence length, 34 phase transition, 17 Phi-value, 150 phosphorylation, 168 pi-calculus, 218 Pochhammer symbol, 138 Poisson distribution, 13 Poisson equation, 159 Poisson process, 43 Poisson-Boltzmann equation, 160 polylog function, 102 pressure, 8 primary structure, 95 prior probability distribution, 3 probability, 2 probe, 117 process, 219 promoter site, 185 protein complex, 95 protein degradation, 46 protein-ligand docking, 156 protein-protein docking, 156 pseudoknot, 89, 92, 129 quaternary structure, 95 rainbow diagram, 89, 92 Random Energy Model, 153 random function, 42 random graph, 229 Random Matrix Theory, 137 random walk, 32 rate constants, 65 rate equation, 66 Rayleigh particle, 58 reaction-diffusion equation, 195 realization, 42 regulatory region, 113 renormalization group, 17, 18 repressilator, 217 reversible, 55 RNA, 83 RNA shape, 145 RNA topology, 134, 135

Index robustness, 234 saddle-point, 21 sample function, 42 scale invariant, 24 scale-free, 230 scaling form, 27 score, 69 scoring function, 243 scoring matrix, 69 second-order phase transition, 25 secondary structure, 86, 95 self-adjoint, 134 self-avoidong walk, 104 self-hybridization, 86 sense strand, 84 sensitivity, 173 separation of variables, 47 sequence alignment, 68 sequence disorder, 140 sequence-dependent, 88 short-range activation, 195 signalling cascade, 210 small-world networks, 241 Smith-Waterman, 72 solvation free energy, 157 specific heat, 26 specific heat exponent, 27 spectral density, 137 spike-in genes, 118 spurious drift term, 62 stability gap, 155 stacking, 98 stacking interactions, 84 standard normal distribution, 11 stationary phase approximation, 21 stationary point, 196 stationary process, 44 statistical ensemble, 1 step operator, 48 stiffness, 33 stochastic delay, 219 stochastic focusing, 217 stochastic reaction rate, 67 stochastic variable, 42

Index structural factor, 152 susceptibility, 26 symmetric random walk, 48 symmetry breaking, 197 systems biology, 185 tail bridging, 171 target RNA, 117 temperature, 8 tertiary contact, 152 tertiary structure, 95 thermal boundary, 115 thermal denaturation, 97 thermal energy, 4 thermodynamic equilibrium, 5 thermodynamic limit, 4 total differential, 7 transcription, 84 transcription factor, 167, 185 transcription factories, 167 transfer matrix, 14 transfer RNA, 86 transition probability, 43 transition state, 150 treelike, 230 Turing instability, 197 Turing pattern, 197 two-state folder, 150 universal form, 27 universality class, 73 upper critical dimension, 31 UV-divergence, 29 Van der Waals radius, 156 variation, 3 vertex, 229 vertex degree, 230 weak-field limit, 27 Wiener-L´evy process, 43 work, 5 Worm-Like Chain (WLC), 34


FIGURE 1: RNA structure predicted by the program Kinefold for a 79 bp mini-RNA of Xenopus tropicalis. Direct output of the program, see calculated free energies at the bottom.

FIGURE 2: Two-state folder CI2, the chymotrypsin inhibitor 2. Structural elements are an α-helix and beta sheets (flat arrows). C. Merlo et al.; Copyright (2005) National Academy of Sciences U.S.A.

FIGURE 3: The folding funnel; a) folding paths from two different initial conditions: A follows a direct path whereas B is trapped in a metastable intermediate; b) rugged energy landscape with multiple minima at intermediate energies. Note that the ‘true’ equilibrium has a much lower free energy. (Reprinted with permission from Macmillan Publishers Ltd., K. A. Dill and H. S. Chan, 1997.)

FIGURE 4: The local electrostatic potential of a microtubule. Red and blue values indicate values of the potential of opposite sign. N. A. Baker et al.; Copyright (2001) National Academy of Sciences U.S.A.

FIGURE 5: Comparison of local (left) and nonlocal (right) electrostatic potential of the enzyme trypsin. The color code indicates postive and negative values of the potential for a selected threshold. Reprinted with permission from A. Hildebrandt, Copyright (2005) Rhombos-Verlag.

FIGURE 6: Chromatin structure. The base level of compaction is the 11 nm fiber, in which the nucleosomes are positioned as ‘beads on a string’; nucleosomal arrays condense to form the 30 nm fiber of yet unknown structure; higher order compactions lead to the familiar chromosome shape of condensed chromatin. (Reprinted with permission from Macmillan Publishers Ltd., G. Felsenfeld and M. Groudine, 2003.)

FIGURE 7: Nucleosomal core particle in a ribbon presentation of the core histones (H2A: green; H2B: light green; H3: cyan; H4: red) bound to DNA (yellow). The axis of the wrapped DNA is represented by a white line following the centerline between the base pairs. Notice the dangling histone tails; the first 20 residues of the H3 tails are omitted for clarity. (Reprinted with permission from Macmillan Publishers Ltd., G. Felsenfeld and M. Groudine, 2003.)

FIGURE 8: a) Protein network graph. Is such a network scale-free? See the range of observed scaling behaviour under b). (Reprinted with permission from Macmillan Publishers Ltd., H. Jeong et al., 2001.)